JP2007079019A - Method and device for speech synthesis, and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To emphasize a keyword part by altering a speech of the keyword part except the speaker, sound volume, pitch, or speaking speed. <P>SOLUTION: A device for speech synthesis is equipped with a cadence prediction information generation section which generates cadence prediction information 502, a scale calculation section which finds scales of respective symbols of a text body by mapping from a feature quantity to a scale based upon the cadende prediction information and the respective symbols of the text body, a phoneme selection section 104 which selects a text phoneme from a phoneme database 109, a keyword scale calculation section 107 which finds a scale based upon the meter prediction information and the respective symbols of keywords in the text body, a keyword priority phoneme selection section 107 which selects the keyword phoneme from a phoneme database, a phoneme replacement section 107 which replaces a phoneme of a keyword part of the text phoneme with the keyword phoneme, and a synthesis section 111 which connects phonemes. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to speech synthesis for reading a text body, and more particularly to a speech synthesizer, a speech synthesis method, and a computer program for emphasizing keywords with good sound quality.

  2. Description of the Related Art Generally, a speech synthesizer is known that converts a text document stored in a PC (personal computer) into a voice and reads it out based on a natural voice of a person recorded in advance. The speech synthesizer synthesizes speech based on a corpus in which natural speech that can be divided into parts of speech is recorded.

  In order to synthesize speech, a conventional speech synthesizer first performs, for example, morpheme analysis and dependency analysis on input text, and converts it into phoneme symbols, accent symbols, and the like.

  Next, the conventional speech synthesizer uses the part of speech information of the input text obtained from the phoneme symbol, the accent symbol string, and the morpheme analysis result, and uses the phoneme duration (voice length), fundamental frequency (voice height). And vowel-centered power (voice volume).

  Next, the conventional speech synthesizer is connected to the synthesis unit (phoneme fragment) that is closest to the estimated phoneme duration, fundamental frequency, vowel center power, etc. and stored in the waveform dictionary. The combination of composition units with the smallest distortion is selected using dynamic programming or the like. The unit selection performed at this time uses a scale that matches the perceptual feature.

  When the combination of the synthesis units is selected, the conventional speech synthesizer synthesizes the speech by connecting the phonemes according to the selected phoneme combination.

  In addition, speech synthesis apparatuses according to the related art include speech synthesis apparatuses that can emphasize important parts in a document and parts that a document creator particularly wants to convey to a reader and read them out (for example, Patent Document 1, Patent Document 1). reference).

JP-A-10-274999

  However, the conventional speech synthesizer emphasizes the keyword by calculating the cost from the whole text body entered and changing either the speaker, volume, pitch, or speaking speed of the keyword part. However, there was no speech synthesizer that changed the others and emphasized the keyword part.

  In addition, in the conventional speech synthesizer, the cost was calculated from the whole text body entered, and the keyword was emphasized because the keyword was changed by changing the speaker, volume, pitch, or speaking speed. It was difficult to improve the sound quality of the keyword part.

  The present invention has been made in view of the above problems, and an object of the present invention is to enhance the keyword portion by changing the speech of the keyword portion other than the speaker, volume, pitch, or speaking speed. A new and improved speech synthesis apparatus, speech synthesis method, and computer program are provided.

  In order to solve the above problems, according to a first aspect of the present invention, a text body is analyzed, and prosodic prediction information is generated by predicting a feature quantity of each symbol (or each phoneme symbol) constituting the text body. The text body is constructed by mapping from the observable feature amount to a scale determined based on a perceptual experiment based on the prosodic prediction information generating unit and each symbol constituting the text body. A scale calculation unit for obtaining a scale of each symbol to be recorded; a phoneme database in which a plurality of voices are recorded and phonemes constituting the speech are stored; and for each symbol constituting the text body, the scale calculation unit calculates A phoneme selection unit that selects text phonemes from a phoneme database based on the measured scale; observable based on prosodic prediction information and each symbol constituting a keyword in the text body A keyword scale calculation unit that calculates a scale of each symbol that constitutes the keyword by mapping from a characteristic amount to a scale that is determined based on a perceptual experiment; and a keyword scale calculation unit that obtains a scale of each symbol that constitutes the keyword; A keyword-priority phoneme selection unit that selects a keyword phoneme from a phoneme database based on a scale of each symbol constituting the selected keyword; and a keyword-priority phoneme selection of a phoneme corresponding to the keyword part among the text phonemes selected by the phoneme selection unit There is provided a speech synthesizer comprising: a phoneme replacement unit that replaces a keyword phoneme selected by the unit; and a synthesis unit that connects phonemes and generates synthesized speech.

  According to the present invention, the phoneme selection unit provided in the speech synthesizer selects the text phoneme, the keyword priority phoneme selection unit selects the keyword phoneme, and the phoneme replacement unit replaces the corresponding part of the text phoneme with the keyword phoneme. According to such a configuration, the keyword scale calculation unit calculates the scale based on each symbol constituting the keyword and the prosodic prediction information separately from the phoneme selection unit selecting the text phoneme. A keyword priority phoneme selection unit selects a keyword phoneme. Therefore, the scale is determined by limiting the target range to the keyword part, and the phoneme is selected, so that the naturalness of the synthesized speech of the keyword part increases, and the synthesized speech of the keyword part becomes clear, and the keyword part is emphasized .

  The text phoneme is, for example, a phoneme selected for each symbol constituting the text body. The keyword phoneme is, for example, a phoneme selected for each symbol constituting the keyword in the text body.

  In addition, the scale calculation unit and the phoneme selection unit according to the present invention are configured as separate bodies. However, the present invention is not limited to this example. For example, the scale calculation unit and the phoneme selection unit may be configured as a single unit. .

  In addition, the keyword scale calculation unit, the keyword priority phoneme selection unit, and the phoneme replacement unit according to the present invention are configured separately, but are not limited to this example. For example, the keyword scale calculation unit and the keyword priority phoneme selection unit And the phoneme replacement unit may be configured integrally.

  The feature quantity is, for example, an acoustic scale, a physical quantity, language information, and the like. The physical quantity is, for example, a spectrum, a mel cepstrum, a pitch, a sound length, etc., but is not limited to such an example. The perceptual evaluation amount (psychological amount) is, for example, a local cost (or local scale, sub cost), but is not limited to such an example. The mapping from the feature quantity to the psychological quantity uses, for example, a cost function, but is not limited to this example.

  When there are two or more keywords in the text body, the keyword priority phoneme selection unit assigns a priority to each of the keywords in the order in which the keywords appear from the top of the text body, The weights assigned to the keywords may be converted into priorities, and the keyword phonemes may be selected in the order of priority.

  When there are two or more keywords in the text body, the keyword priority phoneme selection unit weights each keyword and weights the keywords in the order in which the keywords appear. You may comprise so that the scale of each symbol which comprises may be calculated | required.

  The phoneme replacement unit sequentially replaces the phoneme corresponding to the keyword part in the text phoneme selected by the phoneme selection unit with the keyword phoneme when the keyword phoneme selection unit selects the keyword phoneme. You may comprise.

  The phoneme replacement unit is configured to determine whether or not to replace with the keyword phoneme by comparing the scale obtained by the keyword priority phoneme selection unit with the measure obtained by the phoneme selection unit. Also good.

  The phoneme replacement unit may compare the difference between the scale obtained by the keyword priority phoneme selection unit and the scale obtained by the phoneme selection unit with a predetermined threshold.

  In order to solve the above-mentioned problem, according to another aspect of the present invention, a prosodic prediction information generating step of analyzing a text body and generating prosodic prediction information predicting a feature amount of each symbol constituting the text body; Mapping the scale of each symbol constituting the text body by mapping from the observable feature amount to the scale determined based on the perceptual experiment based on the prosodic prediction information and the symbols constituting the text body; A scale calculation step to be obtained; for each symbol constituting the text body, a text from a phoneme database storing one or more phonemes constituting the recorded speech based on the scale obtained in the scale calculation step A phoneme selection step of selecting a phoneme; the observable based on the prosodic prediction information and each symbol constituting a keyword in the text body A keyword scale calculating step for obtaining a scale of each symbol constituting the keyword by mapping the collected amount to a scale determined based on the perceptual experiment; and the keyword scale calculating step for each symbol constituting the keyword A keyword-preferred phoneme selection step of selecting a keyword phoneme from the phoneme database based on the scale of each symbol constituting the keyword obtained in step; a phoneme corresponding to the keyword portion of the text phoneme selected in the phoneme selection step A speech synthesizing method comprising: a phoneme replacement step that replaces the keyword phoneme with the keyword phoneme selected in the keyword priority phoneme selection step; and a synthesized speech generation step that generates a synthesized speech by connecting the phonemes. The

  In order to solve the above-mentioned problem, according to another aspect of the present invention, prosody prediction information generation that analyzes a text body and generates prosodic prediction information that predicts a feature amount of each symbol constituting the text body is performed. Mapping each of the symbols constituting the text body by mapping from an observable feature amount to a scale determined based on a perceptual experiment based on the prosodic prediction information and the symbols constituting the text body. A scale calculation step for obtaining a scale; a phoneme database in which one or more phonemes constituting the recorded speech are stored for each symbol constituting the text body based on the scale obtained in the scale calculation step A phoneme selection step of selecting a text phoneme from: the observation based on the prosodic prediction information and each symbol constituting a keyword in the text body A keyword scale calculating step for obtaining a scale of each symbol constituting the keyword by mapping from the feature value to a scale determined based on the perceptual experiment; and the keyword scale for each symbol constituting the keyword A keyword-preferred phoneme selection step for selecting a keyword phoneme from the phoneme database based on the scale of each symbol constituting the keyword obtained in the calculation step; and corresponding to the keyword portion of the text phoneme selected in the phoneme selection step Causing a computer to execute a speech synthesis method including a phoneme replacement step of replacing a phoneme to be replaced with the keyword phoneme selected in the keyword priority phoneme selection step; and a synthesized speech generation step of generating the synthesized speech by connecting the phonemes Featuring a computer Program is provided.

  In order to solve the above-described problem, according to another aspect of the present invention, a text body expressed in kanji characters is converted into prosodic prediction information, and an optimal value is obtained for each phoneme unit according to the prosodic prediction information. Select phonetic candidates from corpus (phoneme database) that holds recorded speech by cost calculation (or scale calculation) using a cost function that maps observable features to psychological quantities. Then, by connecting the waveform segments (waveform data) corresponding to the selected phoneme, the speech synthesizer that obtains the synthesized waveform can perform cost recalculation independently for a predetermined specific word (keyword). A speech synthesizer including a keyword-priority phoneme selection unit that re-decides an optimal phoneme candidate is provided.

  In the keyword priority phoneme selection unit, the result of cost recalculation performed independently for a predetermined specific word (keyword) indicates the naturalness of the synthesized sound of each specific word (keyword) part. A combination of phoneme candidates may be selected by comparing the cost values.

  In the keyword priority phoneme selection unit, when two or more predetermined specific words (keywords) exist, the keyword priority phoneme selection unit is operated in the order in which the specific words (keywords) appear from the head of the text body, or Priority is given to a plurality of specific words (keywords), and the keyword priority phoneme selection unit is operated in the priority order, or weights are given to a plurality of specific words (keywords), and the specific word (keyword) appears. The weighted cost recalculation may be performed in order, or the weights assigned to a plurality of specific words (keywords) may be converted into priorities, and the keyword priority phoneme selection unit may be operated in order of priority.

  In the keyword-preferred phoneme selection unit, for the first time, the specific word (keyword) portion of the combination of phoneme candidates selected in the entire text body expressed in kanji characters is sorted in the order in which the specific word (keyword) appears, or , In the order of priority, or in the order that results from the comparison of the cost values recalculated with weighting, the phoneme candidate obtained by applying the keyword priority phoneme selection unit to each specific word (keyword) You may make it replace with a combination.

  In the keyword-preferred phoneme selection unit, for each specific word (keyword), a specific word (keyword) portion of the combination of phoneme candidates selected in the entire text body expressed in Kanji characters for the first time You may comprise so that the keyword priority phoneme selection part which added the means to judge whether to replace with the combination of the phoneme candidate calculated | required by operating the priority phoneme selection part may be provided.

  In the first time, the keyword priority phoneme selection unit is applied to each specific word (keyword) for the specific word (keyword) portion of the combination of phoneme candidates selected in the entire text body expressed in Kanji characters. In the keyword-preferred phoneme selection unit, which adds a means to determine whether or not to replace the phoneme candidate combination obtained by the above, for the specific word (keyword) part in the text body expressed in kanji kana characters, First, obtain the cost (or scale) of the specific word (keyword) part obtained when selecting phoneme candidates for the entire text sentence, and compare it with the cost value obtained by operating the keyword priority phoneme selection unit If a replacement is necessary, the keyword-preferred phoneme selection unit may be replaced with the combination of phoneme candidates obtained by the action. .

  As described above, according to the present invention, the scale of each symbol constituting the keyword in the text body is separately provided independently of obtaining the scale of each symbol constituting the text body and selecting the text phoneme. Then, the keyword phoneme is selected, and the keyword part is replaced with the keyword phoneme to clearly emphasize the keyword part.

  DESCRIPTION OF EXEMPLARY EMBODIMENTS Hereinafter, preferred embodiments of the invention will be described in detail with reference to the accompanying drawings. In the following description and the accompanying drawings, components having substantially the same functions and configurations are denoted by the same reference numerals, and redundant description is omitted.

(About voice synthesizer)
First, the speech synthesis apparatus 100 according to the first embodiment will be described with reference to FIG. FIG. 1 is a block diagram illustrating a schematic configuration of the speech synthesizer according to the first embodiment.

  As shown in FIG. 1, the speech synthesizer 100 includes a text analysis unit 101, a prosody prediction unit 103, a phoneme selection unit 104, a keyword priority phoneme selection unit (or phoneme replacement unit) 107, a corpus (or A phoneme database) 109 and a phoneme connection unit 111.

  Note that the case where the text analysis unit 101 and the prosody prediction unit 103 according to the present embodiment are separated from each other will be described as an example. However, the present invention is not limited to this example. For example, the text analysis unit 101 and the prosody prediction unit 103 The present invention can be implemented even when 103 is integrally configured as a prosody prediction information generation unit.

  As shown in FIG. 1, the text analysis unit 101 converts a text body expressed by kanji characters to phonemic symbols. Note that the phoneme is a unit that can be segmented as represented by a phoneme symbol, but is not limited to such an example.

  More specifically, the text analysis unit 101 performs, for example, a morphological analysis and a dependency analysis on a text body expressed by Kanji characters, and displays an accent symbol string and a morpheme representing part of speech information of the text body. Output analysis results.

The prosody prediction unit 103 includes the phoneme symbol of the text body converted by the text analysis unit 101, the accent symbol string output from the text analysis unit 101, and the text body obtained from the morphological analysis result by the text analysis unit 101. Using the part-of-speech information, the pitch (voice pitch: fundamental frequency F ₀ ), phoneme duration (voice length), and mel cepstrum representing the waveform components are predicted. Details of the mel cepstrum and the like are described in, for example, Japanese Patent Application Laid-Open No. 2003-208188.

  The phoneme selection unit 104 selects a phoneme from the corpus 109 using the pitch predicted by the prosody prediction unit 103, the phoneme duration, and the mel cepstrum as parameters for phoneme selection processing. The corpus 109 is, for example, a phoneme database that records a plurality of voices with different speakers, pitches, etc., and stores at least one or more phonemes constituting the recorded voices, and is a hard disk drive (HDD). ) Or the like.

  In the process of selecting the phoneme, the phoneme selection unit 104 uses a scale (hereinafter, cost) that matches the perceptual characteristic. In addition, the cost function that maps from the observable feature quantity (physical quantity, etc.) to the cost (or perceptual evaluation quantity (psychological quantity)) determined based on the perceptual experiment is, for example, the sub-cost related to prosody and the pitch Consists of a function that is a sum of five weighted subcost functions: subcost for discontinuity, subcost for phonological environment substitution, subcost for spectrum discontinuity, and subcost for phonological suitability. (For example, refer to JP2003-208188).

  The feature quantity is, for example, an acoustic scale, physical quantity, linguistic information, etc., and the physical quantity is, for example, spectrum, mel cepstrum, pitch, length of sound, rhythm (sound pitch, etc.), etc. However, it is not limited to such an example. The psychological quantity is, for example, a local cost (or local scale, sub cost), but is not limited to such an example. For example, a cost function is used for the mapping from the feature quantity to the psychological quantity. Note that the naturalness of the synthesized sound tends to increase as the cost value obtained from the cost function is minimized.

  Further, the cost function according to the present embodiment needs to reflect an auditory impression, and thus is determined based on a perceptual experiment. The above perception experiment is performed because it is difficult to grasp how much the physical quantity changes and the sound quality deteriorates. In the perceptual experiment, for example, the subject listens to the synthesized speech in which the text is read out, and the naturalness is evaluated in 5 levels (for example, 1 level “unnatural” to 5 levels “natural speech and no inferiority”). It is not limited to.

  The phoneme selection unit 104 acquires, using the corpus 109, a subcost value related to phoneme suitability, a subcost value related to phonological environment substitution, and a subcost value related to prosody for the phoneme of the text body. A target phoneme is selected as a candidate from a minimum sub-cost value (hereinafter referred to as a target cost value) obtained by adding the sub-cost values within a range of values having a certain range.

  As shown in FIG. 4, the phoneme selection unit 104 starts from the text body start position (for example, the body start shown in FIG. 4) based on the phoneme candidate obtained by the target selection (for example, the text body end position (for example, Until the end of the text shown in FIG. 4, the combination of the target cost value, the sub-cost value related to the pitch discontinuity, and the sub-cost value related to the spectrum discontinuity is combined into the dynamic phoneme candidate that minimizes the cost value. Select using programming. FIG. 4 is an explanatory diagram showing an example of the outline of the phoneme selection process according to the present embodiment.

  As shown in FIG. 4, first, the sequence of phonological symbols (or symbols) from the start of the body to the end of the body shown in the uppermost row is, for example, a sequence of phonological symbols output by the text analysis unit 101. 4 is a sequence of phonemes (or text phonemes) selected from the corpus 109 for which the cost is determined by the phoneme selection unit 104, for example. The column shown in the lower row is a column of phonemes (or keyword phonemes) selected from the corpus 109 for which the keyword priority cost is determined by the keyword priority phoneme selection unit 107, and the lower row in FIG. For example, among the phonemes selected by the phoneme selection unit 104 (phonemes in the second row from the top in FIG. 4), the keyword-part phoneme is converted by the keyword priority phoneme selection unit 107. Is a sequence of which is substituted in-option has been phoneme phoneme.

  The phoneme selection unit according to the related art stores a speech waveform database including speech segment data of speech waveform signals corresponding to phoneme labels, as described in, for example, Japanese Patent Application Laid-Open No. 2001-100775. When synthesizing speech by connecting arbitrary phoneme strings by concatenating speech segments of waveform signals, the phoneme selection unit generates a list of a pair of phonemes included in the speech waveform database, and based on the phoneme list A list of pairs of phonemes to be written is generated, and the similarity between the prosodic feature parameter and the acoustic feature parameter for the pair of phonemes to be written is calculated based on the phoneme list, and the similarity is equal to or greater than a predetermined threshold. The speech segment data of the speech waveform signal associated with one of the pair of phonemes is converted into a speech wave. We had to reduce the amount of audio data is deleted from the database.

  The keyword priority phoneme selection unit 107, which is a characteristic part of the speech synthesizer 100, uses the pitch, phoneme duration, and mel cepstrum predicted by the prosody prediction unit 103 as parameters of phoneme selection processing from the corpus 109. Select.

  In the process of selecting a phoneme by the keyword priority phoneme selection unit 107, a scale (or cost) that matches the perceptual characteristic is used. A cost function that maps observable features to psychological quantities is a sub-cost related to prosody, a sub-cost related to pitch discontinuity, a sub-cost related to phonological environment substitution, a sub-cost related to spectral discontinuity, and a phoneme fit. It is constructed as a function that is a sum of five weighted subcost functions with respect to the subcost. Details are described in, for example, Japanese Patent Application Laid-Open No. 2003-208188.

  The keyword priority phoneme selection unit 107 uses the corpus 109 to acquire a subcost value related to phoneme suitability, a subcost value related to phonological environment substitution, and a subcost value related to prosody for the phoneme of the text body. A target phoneme is selected as a candidate from a minimum sub cost value (target cost value) obtained by adding two sub cost values within a range of values having a certain range.

  As shown in FIG. 4, the keyword priority phoneme selection unit 107 starts from the keyword start position specified by the user (for example, the keyword start shown in FIG. 4) in the text body based on the phoneme candidate obtained by target selection. In the range up to the keyword end position (for example, the keyword end shown in FIG. 4), the target cost value, the sub-cost value related to the pitch discontinuity, and the sub-cost value related to the spectrum discontinuity are added together to minimize the cost value. Is selected using dynamic programming.

  The keyword start position and the keyword end position are specified by the number of phonemes from the start position of the text body, for example. For example, as shown in FIG. 4, when the keyword (or keyword phoneme string) is “aoki”, the keyword starting position “a” is specified by the number of phonemes from the starting position “h” of the text body. Can do. The number of phonemes can be specified in the same manner for “o”, “k”, and “i” in order.

  Also, the keyword position can be specified by tagging the keyword part in the text body, analyzing the text body once, displaying the phoneme string of the text body to the user, and then displaying the keyword part as text. There are a method of designating from the text, and a method of storing a plurality of keywords in advance, selecting a keyword from among the plurality of keywords, and searching for a location that matches the selected keyword in the text body.

  In addition, as shown in FIG. 4, when there are two or more keywords, the keyword-priority phoneme selection unit 107 calculates the cost value for the keyword portion from the start of the text to the end of the text shown in FIG. However, the present invention is not limited to such an example. For example, the calculation of the cost value related to the keyword part may be performed when a priority is assigned to a plurality of keywords and the cost value is calculated according to the priority order, or , When assigning weights to multiple keywords and recalculating weighted costs in the order in which the keywords appear, or by converting the weights assigned to multiple keywords into priorities and calculating cost values according to the priorities It is possible to implement even when performing.

  The keyword priority phoneme selection unit 107 calculates weights and cost values in the order in which the keywords appear, in the priority order, or in the order in which the keywords appear, with respect to the keyword portion of the phoneme candidate combination selected in the entire text body. Find the multiplied weighted cost value.

  Next, the keyword priority phoneme selection unit 107 compares the obtained weighted cost values and replaces them with combinations of phoneme candidates selected by keyword priority phonemes in ascending order of the weighted cost values.

  The phoneme connection unit (or synthesis unit) 111 connects, for example, the selected phonemes (or waveform segments) in order from the start position of the text body using the cross-correlation method and the interpolation method. Generate speech (or synthesized waveform). When the synthesized speech is output to the outside, the user can hear the text body by ear.

  Note that the speech synthesizer 100 is a device capable of outputting a synthesized speech based on a text body and a keyword, and can output the text body as speech by outputting the synthesized speech. Device. More specifically, the speech synthesizer 100 includes, for example, an input unit (not shown) corresponding to a CPU, a memory, an HDD (hard disk drive), a mouse, and a display unit (not shown) corresponding to a liquid crystal display. Etc.) can be exemplified, but is not limited to such an example.

  In addition, the display part with which the speech synthesizer 100 concerning this Embodiment is provided outputs the display screen data and audio | voice data processed so that display was possible by CPU. In addition, the display unit is exemplified by a TV or a liquid crystal display device, for example, and both of them are provided with a speaker, and can output a sound or a moving image in addition to a still image.

  The input unit is, for example, a pointing device such as a mouse, a trackball, a trackpad, a stylus pen, a touch panel, or a joystick that can receive an operation instruction from a user, or an operation means such as a keyboard, button, switch, or lever. And an input control circuit for generating an input signal and outputting it to the CPU.

  The user of the speech synthesizer 100 operates the input unit to input various data to the speech synthesizer 100, select keywords, and instruct various processing operations such as output of synthesized speech. can do.

  Note that a conventional speech synthesizer includes, for example, means for storing speech segments, means for extracting and outputting prosodic feature parameters, and the like, as described in Japanese Patent Laid-Open No. 10-049193. A target cost for a phoneme sequence of an input natural utterance sentence, a means for storing a parameter, a means for determining a weighting coefficient vector representing a contribution in an acoustic feature parameter, a means for storing a weighting coefficient vector, And speech unit selection means for searching for a combination of phoneme candidates having the minimum cost including the connection cost and outputting index information, and means for synthesizing and outputting the input speech.

In addition, a conventional speech synthesizer improved from the conventional speech synthesizer so that the sound quality is further improved is, for example, a language processing unit and a prosodic pattern generation unit as described in JP-A-2003-208188. A phoneme unit selection unit, a waveform dictionary, and a speech waveform generation unit. The language processing unit performs morpheme analysis and dependency analysis, and converts the text into phoneme symbols, accent symbols, and the like. The prosodic pattern generation unit uses the part-of-speech information of the input text obtained from the phoneme symbol, the accent symbol string, and the morpheme analysis result, the phoneme duration (voice length), the fundamental frequency (voice pitch F ₀ ), Estimate vowel center power (voice volume). Furthermore, the phoneme unit selection unit uses a cost function that uses physical quantities and psychological quantities directly obtained from the perceptual experiment results. The cost function has five sub-cost functions: a sub-cost related to prosody, a sub-cost related to pitch discontinuity, a sub-cost related to phoneme environment substitution, a sub-cost related to spectrum discontinuity, and a sub-cost related to phoneme (or phoneme) suitability. Consists of

  Further, a conventional speech synthesizer that reads out text, as described in Japanese Patent Laid-Open No. 10-274999, for example, obtains the cost of a keyword from the entire text, and includes the voice speaker, volume, pitch, and speech speed. By changing at least one arbitrarily selected from the above, an audio signal is generated so that the output audio for the keyword portion and the output audio for the other portion are identified.

(Speech synthesis method)
As shown in FIG. 2, first, when the user operates the input unit, for example, the text body including one or more keywords to be emphasized, the start position of the keyword to be emphasized, and the end of the keyword. The position is input (S201). However, the present invention is not limited to this example. For example, the user may specify the entire range from the keyword start position to the keyword end position.

  The keyword start position and keyword end position input in step S201 are specified based on, for example, the number of phonemes from the start position of the text body, but are not limited to this example.

  In step S201, the text body including one or more keywords to be emphasized specified by the user is input to the text analysis unit 101 shown in FIG.

  As shown in FIG. 2, for example, the text analysis unit 101 performs text analysis on a text body expressed by kanji characters and converts them into phonemic symbols (S203).

  The text analysis unit 101 performs text analysis on the text body. Examples of the text analysis include morphological analysis and dependency analysis. The text analysis unit 101 performs morphological analysis and dependency analysis on the text body, and outputs an accent symbol string and a morphological analysis result representing part-of-speech information of the text body.

  In the speech synthesis method according to the first embodiment, an output result having information on the phoneme symbol converted from the text body and the morpheme analysis result is shown in FIG. Thus, although it is set as a text body intermediate language, it is not limited to such an example.

  Here, the phoneme symbol 501 converted from the text body (text body phoneme symbol shown in FIG. 3A) is, for example, “hajime ... oki ...” as shown in FIG. 3A.

  Next, the text body intermediate language output by the text analysis unit 101 is input to the prosody prediction unit 103 shown in FIG.

As shown in FIG. 2, the prosody prediction unit 103 includes a pitch (voice pitch, fundamental frequency F ₀ ), phoneme duration (voice length), fundamental frequency (voice pitch F ₀ ), A mel cepstrum expressing the waveform component is predicted (S207).

  When the pitch, phoneme duration, fundamental frequency, and mel cepstrum are predicted by the prosody prediction unit 103 in step S207, prosody prediction information having such predicted information is input to the phoneme selection unit 104.

  The prosody prediction information is information for predicting the feature quantity of each phoneme symbol by, for example, the pitch, phoneme duration, fundamental frequency, or mel cepstrum, but is not limited to this example.

  Here, in order to explain the prosodic prediction information, referring to FIG. 3A, the prosodic prediction information 502 is obtained for each phoneme symbol 501 ("hajime ... oki ..." shown in the vertical direction of the prosodic prediction information 502 shown in FIG. 3A). "For each phoneme symbol"), "start" indicating the start time of the phoneme, "duration" indicating the duration of the phoneme, "pitch" indicating one or more pitches of the phoneme, and 1 or 2 of the phoneme “Mel cep” representing at least two mel cepstrums.

  As shown in FIG. 2, the phoneme selection unit 104 selects an optimal phoneme using the corpus 109 from the text start position to the text end position of the text body (S214).

  More specifically, step S214 will be described. The phoneme selection unit 104 calculates a subcost value related to phoneme suitability, a subcost value related to phonemic environment substitution, and a subcost value related to prosody for the phoneme of the text body. The phoneme included in the range of values having a certain range is selected as a candidate from the minimum subcost value (target cost value) obtained by adding the above three subcost values.

  As shown in FIG. 4, in the phoneme selection unit 104, the target cost value and the pitch discontinuity from the body start position of the text body to the body end position of the text body based on the phoneme candidates obtained by the target selection. The combination of phoneme candidates that minimizes the sum of the sub-cost value related to spectrum and the sub-cost value related to spectrum discontinuity is selected using dynamic programming. Note that, as the cost value obtained from the cost function is minimized, the naturalness of the synthesized sound is higher. However, the present invention is not limited to this example.

  When the phoneme selection unit 104 completes the phoneme selection of the entire text body (S214), the process proceeds to the processing by the keyword priority phoneme selection unit 107 in order to select the optimum phoneme of the keyword portion (S216, S219). .

  Here, referring to FIG. 3A, a combination 503 of phoneme candidates for the entire text body shown in FIG. 3A includes each phoneme symbol of the text body phoneme symbol 501 (“hajime ... oki ...” in the example shown in FIG. 3A). Information indicating which phoneme in the corpus 109 the corresponding phoneme is in which position from the head is located is described. The information described in the phoneme candidate combination 503 of the entire text body is, for example, “the phoneme corresponding to the phoneme symbol“ h ”is the phoneme existing at the third position from the beginning of one speech file”. Show.

  Further, as shown in FIG. 3B, the keyword priority phoneme selection processing unit 107 includes a prosody prediction information 502, a combination of phoneme candidates 503 of the entire text body output from the phoneme selection unit 104, and a keyword position specified by the user. Are entered.

  As shown in FIG. 2, the keyword priority phoneme selection unit 107 first determines whether the keyword appears last in the text body (S216). Note that the processing in step S501 shown in FIG. 5 described later and the processing in step S216 are substantially the same.

  The keyword priority phoneme selection unit 107 is arranged in the order of keywords appearing from the text start position of the text body, or in the order of priority based on the priority given to the keywords of the text body, or one or more in the text body. The weights assigned to the keywords are converted into priorities, and keyword priority phoneme selection processing is performed in order of priority. For that purpose, it is necessary to determine whether or not it is the last keyword in step S216 shown in FIG.

  If the result of determination in step S216 is not the last keyword, the keyword priority phoneme selection process shown in FIG. 2 is performed.

  On the other hand, if the result of determination in step S216 is the last keyword, the phoneme selection of the keyword portion is completed and the combination of phoneme candidates from which waveform segments can be obtained is determined. The process moves to 111.

  Next, as shown in FIG. 2, the keyword priority phoneme selection unit 107 selects an optimal phoneme from the keyword part in the text body using the corpus 109 (S219).

  Here, the keyword priority phoneme selection process (S219) by the keyword priority phoneme selection unit 107 will be described with reference to FIG. FIG. 5 is a flowchart showing an outline of the keyword priority phoneme selection process according to the first embodiment.

  The keyword priority phoneme selection process (S219) shown in FIG. 2 includes, as shown in FIG. 5, a process for confirming whether or not the keyword is the last keyword (S501), a process for selecting a target for the keyword (S505), A process for setting a combination of phoneme candidates that minimizes the cost value for a keyword (S509) and a process for replacing a combination of phoneme candidates for the entire text body with a combination of phoneme candidates at a keyword location (S513) are included. ing.

  In the process of performing target selection for the keyword (S505), the sub-cost value related to phoneme compatibility, the sub-cost value related to phonological environment substitution, and the sub-cost value related to prosody for the phoneme of the keyword part are stored in the corpus 109. The phoneme is selected as a candidate from the minimum sub-cost value (target cost value) obtained by using the three sub-cost values and included in a range having a certain range.

  In the process of setting a combination of phoneme candidates that minimizes the cost value for the keyword (S509), as shown in FIG. 4, the user in the text body based on the phoneme candidates obtained by target selection. From the specified keyword start position to the keyword end position, the target cost value, the sub cost value related to the pitch discontinuity, and the sub cost value related to the spectrum discontinuity are combined to obtain a phoneme candidate that minimizes the cost value. , Select using dynamic programming.

  The keyword start position and the keyword end position are specified by, for example, the number of phonemes counted from the text start position of the text body. However, the present invention is not limited to this example. For example, the keyword start position and the keyword end position are specified by the number of phonemes counted from the text end position. Etc.

  The keyword position can be specified by, for example, tagging the keyword part in the text body, analyzing the text body once, displaying the phoneme string in the text body to the user, and specifying the keyword part, Alternatively, there is a method in which one or two or more keywords are stored in advance and the user selects a keyword and a search is made for a portion that matches in the text body. However, the method is not limited to this example.

  As shown in FIG. 4, when there are two or more keyword phoneme candidates, the cost value for the keyword portion is calculated in the order of appearance of the keywords, for example, from the text start position to the text end position of the text body.

  Note that the calculation of the cost value related to the keyword part according to the present embodiment is not limited to the case of performing the keyword appearance order from the text start position to the text end position. The present invention can be carried out even when priority is given to a plurality of keywords and the priority order is given.

  As shown in FIG. 5, in the replacement processing (S513) of the combination of phoneme candidates of the whole text body with the combination of phoneme candidates of the keyword portion, the combination of the phoneme candidates of the entire text body of FIG. The phoneme candidate combinations selected in the whole text are replaced with the phoneme candidate combinations selected in keyword order in the order in which the keywords appear or in the priority order with respect to the keyword part. As shown in FIG. 5, step S513 is executed sequentially each time the phoneme selection process of step S214 shown in FIG. 2 is performed, and phoneme replacement is performed.

  After the replacement of the keyword part phoneme candidate combination is completed (S513), the process proceeds to step S501, which is substantially the same as step S216 shown in FIG. 2, and the process is repeated until the last keyword is obtained ( S505 to S513).

  Further, a modified example of the keyword priority phoneme selection process according to the first embodiment will be described with reference to FIG. FIG. 6 is a flowchart showing an outline of the keyword priority phoneme selection process according to the first embodiment.

  As shown in FIG. 6, the keyword priority phoneme selection process includes a process for determining whether or not the keyword is the last keyword (S502), a process for selecting a target for the keyword (S505), and a cost value is minimum. And a process for obtaining a weighted cost value (S510), and a process for replacing a combination of phoneme candidates in the entire text body with a combination of phoneme candidates at a keyword location (S521).

  As shown in FIG. 6, in the process (S510) of setting a phoneme candidate combination that minimizes the cost value for a keyword and obtaining a weighted cost value obtained by multiplying the cost value by the weight, as described above. A combination of phoneme candidates having a minimum cost value is set for the keyword. Further, in the calculation of the cost value related to the keyword portion (S510), a process of assigning weights to one or more keywords and multiplying the weights and the cost values in the order in which the keywords appear is performed. Alternatively, in the calculation of the cost value relating to the keyword portion (S510), the weights assigned to one or more keywords are converted into priorities and performed in order of priority.

  Next, as shown in FIG. 6, it is determined whether or not it is the last keyword, and if it is the last keyword as a result of the determination (S502), in order of keywords having the smallest value of the weighted cost value in FIG. The processing shifts to the replacement processing (S521) of the combination of phoneme candidates in the keyword portion of the combination of phoneme candidates in the entire text body.

  As shown in FIG. 6, in the replacement process (S521) of the combination of phoneme candidates in the keyword text in the combination of phoneme candidates in the entire text body (S521), in the order in which the keywords appear in the combination of phoneme candidates selected in the entire text body. A weighted cost value obtained by multiplying the weight and the cost value is obtained, compared with each other, and replaced with a combination of phoneme candidates selected by keyword priority phoneme in ascending order of the weighted cost value.

  After the replacement of the combination of the phoneme candidates in the keyword part is completed (S521), the process proceeds to the process executed by the phoneme connection unit 111 shown in FIG.

  Returning to FIG. 2 again, as shown in FIG. 2, in the process of connecting waveform segments executed by the phoneme connection unit 111 (S223), the currently handled waveform segment is connected to the next waveform segment.

  When the connection of all the waveform segments is completed by the phonological connection unit 111 (S223), as shown in FIG. 2, the waveform segments connected by the phonological connection unit 111 are output as synthesized speech (or synthesized speech). (S225). By outputting the synthesized sound from an output unit such as a speaker, the speech synthesizer 100 can read out the text while emphasizing the keyword portion in the text body.

This is the end of the description of the series of processes of the speech synthesis method performed by the speech synthesis apparatus 100 according to the first embodiment. The speech synthesis method has the following excellent effects.
(1) A user (user) designates in advance a keyword position to be emphasized in a text body to the speech synthesizer 100, so that the speech synthesizer 100 reads out a natural part with respect to reading a keyword part (natural Reading), the quality of the output synthesized sound is good, and the reading of keywords is emphasized.

(About the second embodiment)
Next, a speech synthesis method according to the second embodiment will be described with reference to FIG. In the speech synthesis method according to the second embodiment described below, differences from the speech synthesis method according to the first embodiment described above will be described in detail, and the other points are substantially the same. The explanation is omitted as there is.

FIG. 7 is a flowchart showing an outline of keyword priority phoneme selection processing according to the second embodiment. The speech synthesis method according to the second embodiment is different in that the keyword priority phoneme selection process in the speech synthesis method according to the first embodiment shown in FIG. 2 is improved. Hereinafter, although described with reference to FIG. 7, in the keyword priority phoneme selection process according to the second embodiment shown in FIG. 7, the cost value for the keyword is C _{key as} shown in step S511 or the like.

As shown in step S515 of FIG. 7, for the keyword part in the text body, the sub-cost value related to phoneme compatibility, the sub-cost value related to phonological environment substitution, the sub-cost value related to prosody, obtained by the phoneme selection unit 104, Is obtained using the corpus 109, and the minimum sub-cost value (target cost value), the sub-cost value related to the pitch discontinuity, and the sub-cost value related to the spectral discontinuity are obtained by adding the above three sub-cost values. The added cost value C _orig is acquired.

In the keyword priority phoneme selection process shown in FIG. 7, as shown in step S517, the difference between the cost value C _key for the keyword and the cost value C _orig obtained by the phoneme selection unit 104 in the keyword part in the text body is _calculated. The difference between the difference value and the threshold (Threshold) is determined. It should be noted that the threshold value according to the present embodiment can be changed as appropriate according to the use environment of the speech synthesis method.

(Speech synthesis method)
As shown in FIG. 2, in the speech synthesis method according to the second embodiment, first, like the speech synthesis method according to the first embodiment, for example, one or more keywords to be emphasized are included. The user inputs the text body, the start position of the keyword to be emphasized, and the end position of the keyword (S201). However, the present invention is not limited to this example. For example, the user may specify the entire range from the keyword start position to the keyword end position.

  The keyword start position and keyword end position input in step S201 are specified based on, for example, the number of phonemes from the start position of the text body, but are not limited to this example.

  In step S201, the text body including one or more keywords to be emphasized specified by the user is input to the text analysis unit 101 shown in FIG.

  Next, in the speech synthesis method according to the second embodiment, as shown in FIG. 2, the text analysis unit 101 performs text analysis on a text body expressed by, for example, kanji kana characters and converts them into phonetic symbols. (S203).

  The text analysis unit 101 performs, for example, morphological analysis and dependency analysis on the text body expressed by the above-described kanji characters, and outputs an accent symbol string and a morphological analysis result representing the part of speech information of the text body. .

  In the speech synthesis method according to the second embodiment, an output result having information on the phoneme symbol converted from the text body and the morpheme analysis result is obtained as the text body intermediate language. However, it is not limited to such an example.

  Here, as explained in the speech synthesis method according to the first embodiment, the phoneme symbol 501 (the text body phoneme symbol shown in FIG. 3A) obtained by converting the text body is, for example, as shown in FIG. "Hajime ... oki ..."

  Next, the text body intermediate language output by the text analysis unit 101 is input to the prosody prediction unit 103 shown in FIG.

As shown in FIG. 2, the prosody prediction unit 103 includes a pitch (voice pitch, fundamental frequency F ₀ ), phoneme duration (voice length), fundamental frequency (voice pitch F ₀ ), A mel cepstrum expressing the waveform component is predicted (S207).

  Prosody prediction information having information (for example, pitch, phoneme duration, fundamental frequency, mel cepstrum) predicted by the prosody prediction unit 103 (S207) is input to the phoneme selection unit 104.

  Here, as explained in the speech synthesis method according to the first embodiment, referring to FIG. 3A, the prosody prediction information 502 is obtained for each phoneme symbol of the phoneme symbol 501 (the vertical direction of the prosody prediction information 502 shown in FIG. 3A). “Start” indicating the phoneme start time, “duration” indicating the duration of the phoneme, and one or more pitches of the phoneme. “Pitch” and at least “Mel cep” representing one or more mel cepstrums of phonemes.

  As shown in FIG. 2, the phoneme selection unit 104 selects an optimal phoneme using the corpus 109 from the text start position to the text end position of the text body (S214).

  Step S214 has been described in the first embodiment, but more specifically, the phoneme selection unit 104 has a subcost value related to phoneme suitability and a subcost related to phoneme environment substitution for the phoneme of the text body. The value and the sub cost value related to the prosody are acquired using the corpus 109, and within the range of values having a certain range from the minimum sub cost value (target cost value) obtained by adding the above three sub cost values. The phonemes included in the range (for example, within the range of the value obtained by adding ± 5 to the target cost value) are selected as candidates.

  As shown in FIG. 4, in the phoneme selection unit 104, based on the phoneme candidates obtained by target selection, the target cost value and the pitch of the range from the text body start position to the text body end position are determined. A combination of phoneme candidates that minimizes the sum of the sub-cost value related to the discontinuity and the sub-cost value related to the spectrum discontinuity is selected using dynamic programming. Note that the naturalness of the synthesized sound tends to increase as the cost value obtained from the cost function is minimized.

  When the phoneme selection of the entire text body is completed (S214), the process proceeds to the process by the keyword priority phoneme selection unit 107 in order to select the optimum phoneme of the keyword part (S216, S219).

  As shown in FIG. 3A, the phoneme candidate combination 503 of the entire text body includes a phoneme corresponding to each phoneme symbol of the text body phoneme symbol 501 ("hajime ... oki ..." in the example shown in FIG. 3A). Information indicating which sound file exists in 109 and which phoneme is located from the beginning is described.

  Further, as shown in FIG. 3B, the keyword priority phoneme selection processing unit 107 includes a prosody prediction information 502, a combination of phoneme candidates 503 of the entire text body output from the phoneme selection unit 104, and a keyword position specified by the user. Are entered.

  As shown in FIG. 2, as described in the first embodiment, the keyword priority phoneme selection unit 107 determines whether or not the keyword appears last in the text body (S216).

  The keyword priority phoneme selection unit 107 performs processing in the order of keywords appearing from the text start position of the text body, or performs processing in order of priority based on the priority assigned to the keywords of the text body, or in the text body It is necessary to convert the weights assigned to one or more of the keywords into priority levels and perform processing in the priority order. For that purpose, it is necessary to determine whether or not it is the last keyword in step S216 shown in FIG.

  If the result of determination in step S216 is not the last keyword, the keyword priority phoneme selection process shown in FIG. 2 is performed.

  On the other hand, if the result of determination in step S216 is the last keyword, the phoneme selection of the keyword portion is completed and the combination of phoneme candidates from which waveform segments can be obtained is determined. The process moves to 111.

  Next, as shown in FIG. 2, the keyword priority phoneme selection process (S219) is a process for selecting an optimal phoneme from the keyword part in the text body using the corpus 109.

  Here, the keyword priority phoneme process (S219) according to the second embodiment will be described with reference to FIG. FIG. 7 is a flowchart showing an outline of keyword priority phoneme selection processing according to the second embodiment.

As shown in FIG. 7, the keyword priority phoneme selection process (S219) includes a process for confirming whether the keyword is the last keyword (S501), a process for selecting a target for the keyword (S505), and a keyword. A process of setting a combination of phoneme candidates that minimizes the cost value C _key (S511), a process of acquiring the cost value C _orig obtained by the phoneme selection unit 104 at the keyword part in the text body (S515), and C _key The process of verifying whether or not the relationship of -C _orig > threshold is established (S517) and the process of replacing the combination of phoneme candidates in the keyword part of the combination of phoneme candidates in the entire text body (S513) are performed. Is called. Note that the process of confirming whether it is the last keyword (S501) is substantially the same as step S216 described above.

  As shown in FIG. 7, in the process of performing target selection for a keyword (S505), the sub-cost value related to phoneme compatibility, the sub-cost value related to phoneme environment substitution, and the sub-cost related to prosody for the phoneme of the keyword part. Values obtained using the corpus 109, and the phonemes included in the range of values having a certain range from the minimum sub cost value (target cost value) obtained by adding the above three sub cost values. Select a target as a candidate.

In the process of setting a combination of phoneme candidates that minimizes the cost value C _key for the keyword shown in FIG. 7 (S511), as shown in FIG. 4, the phoneme candidate (or phoneme selection) obtained by target selection is set. Originally, the target cost value, sub-cost value related to pitch discontinuity, and sub-cost value related to spectrum discontinuity are added from the keyword start position specified by the user to the keyword end position in the text body. The phoneme candidate combination with the lowest cost value is selected using dynamic programming.

  The keyword start position and the keyword end position can be specified by, for example, the number of phonemes counted from the text start position of the text body, but not limited to such an example, for example, specified by the number of phonemes counted from the text end position It may be the case.

  The keyword position (keyword start position to keyword end position) can be specified by tagging the keyword part in the text body or by analyzing the text body once and displaying the phoneme string in the text body to the user. There are a method of designating a keyword portion and a method of holding one or more keywords in advance, selecting a keyword by the user, and searching for a location that matches the selected keyword in the text body.

  As shown in FIG. 4, when there are two or more keyword phoneme candidates, the cost value for the keyword part is calculated in the order of appearance of the keywords from the text start position to the text end position of the text body, for example. Alternatively, the calculation of the cost value for the keyword part is performed in order of priority by giving priority to one or more keywords, or the calculation of the cost value for the keyword part is 1 or 2 as shown in FIG. Weights are assigned to the above keywords, and the process of multiplying the weights by the cost values is performed in the order in which the keywords appear. Alternatively, the cost values for the keyword part are calculated by assigning the weights assigned to multiple keywords to the priority order. And is performed in order of priority.

As shown in FIG. 7, in the process of obtaining the cost value C _orig obtained by the phoneme selection unit 104 for the keyword portion in the text body (S515), the phoneme selection unit 104 obtains the keyword portion in the text body. Further, a sub cost value related to phonological compatibility, a sub cost value related to phonological environment substitution, and a sub cost value related to prosody are acquired using the corpus 109, and the minimum sub cost value (target) is obtained by adding the above three sub cost values. (Cost value), a sub cost value related to pitch discontinuity, and a sub cost value related to spectrum discontinuity are added to obtain a cost value C _orig .

The cost value C _orig is a cost value obtained based on the sub cost value obtained from the range of the keyword part in the text body, whereas the cost value C _key is determined from the body start position of the text body. This is the cost value of the keyword portion obtained based on the sub cost value obtained from the range up to the end position. That is, both the cost value C _orig and the cost value C _key have different phoneme target ranges for obtaining the three sub cost values, for example. The cost value C _orig tends to be directly expressed as numerical values such as voice quality that one or more phonemes constituting the keyword part originally have.

As shown in FIG. 7, in the process of determining whether the relationship of C _key -C _orig > threshold is satisfied (S517), the phonological selection unit 104 obtains the cost value for the keyword using C _key and the keyword part in the text body. A difference from the cost value C _orig is taken, and a certain threshold value (Threshold) is determined.

  As a result of the above determination, if the difference value is larger than a certain threshold (Threshold) and the above inequality relationship is maintained, the combination of phoneme candidates of the entire text body shown in FIG. Processing shifts to combination replacement (S513).

  When the obtained difference value is equal to or less than a certain threshold value (Threshold), a process of determining whether or not it is the last keyword shown in FIG. 7 is performed (S501).

  As shown in FIG. 7, in the replacement of the combination of phoneme candidates in the keyword text with the combination of phoneme candidates in the entire text body (S513), the keyword for the keyword part of the combination of phoneme candidates selected in the entire text body is used. The priority phoneme selection unit 107 replaces the combination with the phoneme candidate selected by the keyword priority phoneme. In other words, the phoneme at the keyword location is the phoneme selected by the keyword priority phoneme by the keyword priority phoneme selection unit 107, and the phoneme is output as a synthesized sound to be described later.

  In the speech synthesis method according to the second embodiment, the threshold is set so that the phoneme selected by the keyword priority phoneme selection unit 107 is read out more clearly and clearly when it is output as a synthesized speech. The replacement process (S513) is performed.

  After the replacement of the keyword part phoneme candidate combination is completed (S513), the process proceeds to step S501, which is substantially the same as step S216 shown in FIG. 2, and the process is repeated until the last keyword is obtained ( S505 to S513).

  Here, a modified example of the keyword priority phoneme selection process according to the second embodiment will be described with reference to FIG. FIG. 8 is a flowchart showing an outline of keyword priority phoneme selection processing according to the second embodiment.

As shown in FIG. 8, the keyword priority phoneme selection process (S219) includes a process for determining whether or not it is the last keyword (S501), a process for selecting a target for the keyword (S509), a keyword, Processing for setting a combination of phoneme candidates that minimizes the cost value C _key (S511), and obtaining C _orig by multiplying the cost value obtained by the phoneme selection unit 104 in the keyword part in the text body with a weight. processing (S516), processing for determining whether or not the last keyword _(S502), and determines the processing or the relationship of the C _key -C orig> threshold and (S517), a combination of text body entire phoneme candidate , Processing for replacing with a combination of phoneme candidates at the keyword portion (S513).

  As shown in FIG. 8, the keyword-priority phoneme selection process (S219) according to the second embodiment is different from the keyword-priority phoneme selection process according to the second embodiment shown in FIG. Instead, it differs in that it includes step S516 and further includes step S502.

In step S516 in FIG. 8, the sub-cost value related to phoneme compatibility, the sub-cost value related to phonological environment substitution, and the sub-cost value related to prosody obtained by the phoneme selection unit 104 for the keyword portion in the text body are corpus. 109, and the minimum sub-cost value (target cost value), the sub-cost value related to the pitch discontinuity, and the sub-cost value related to the spectral discontinuity are added together. The cost value C _orig is acquired.

  Further, in step S516 in FIG. 8, a weighted cost value obtained by multiplying the weight and the cost value obtained above is obtained in the order in which the keywords appear, and each is compared to obtain a weighted cost value.

Further, as shown in FIG. 8, in the process of acquiring _Corig obtained by multiplying the cost value obtained by the phoneme selection unit 104 by the weight in the keyword part in the text body (S516), the keyword part in the text body is obtained. Thus, the sub-cost value related to the phoneme compatibility, the sub-cost value related to the phoneme environment substitution, and the sub-cost value related to the prosody obtained by the phoneme selection unit 104 are acquired using a corpus, and the three sub-cost values are added together. The cost value C _orig obtained by adding the minimum sub cost value (target cost value), the sub cost value related to the pitch discontinuity, and the sub cost value related to the spectrum discontinuity is obtained, and the cost value C _orig and a plurality of keywords are obtained. Multiply the weight _assigned to the in the order in which the keywords appear, and obtain the weighted cost value C _orig The

As shown in FIG. 8, in the process of determining whether or not it is the last keyword (S502), for example, the weighted cost value C _orig obtained for each keyword is compared, and a small weighted cost value is held. It is determined whether or not the keyword appears last in the text body in the order in which the keywords are displayed.

As a result of the determination (S502), if it is not the last keyword, the process proceeds to processing (S517) for verifying the relationship of “C _key −C _orig > threshold” shown in FIG.

  On the other hand, as a result of the determination (S502), if it is the last keyword, the phoneme selection of the keyword part is completed, and the combination of phoneme candidates from which waveform segments can be obtained is determined. Transition to processing.

  As described in the speech synthesis method according to the first embodiment, as shown in FIG. 2, the process of connecting waveform segments executed by the phoneme connection unit 111 (S223) is the waveform segment currently handled. To the next waveform segment.

  When the connection of all the waveform segments is completed by the phonological connection unit 111 (S223), as shown in FIG. 2, the waveform segments connected by the phonological connection unit 111 are output as synthesized speech (or synthesized speech). (S225). By outputting the synthesized sound from an output unit such as a speaker, the speech synthesizer 100 can read out the text while emphasizing the keyword.

Although the description of the speech synthesis method according to the second embodiment is finished as described above, the speech synthesis method has the following excellent effects.
(1) The keyword portion to be emphasized in the text body is not replaced with the phoneme candidate selected by the keyword priority phoneme selection unit 107 without any evaluation, and the phoneme selection unit 104 is applied to the keyword portion. in determined and compared with the cost value C _key and the keyword priority phoneme selector 107 cost value C _orig obtained, upon replacement necessary, by replacing the combination of phonological candidate keyword priority phoneme unit 107 selects, The keyword part can be read more naturally, the sound quality is good, and the keyword reading can be emphasized.

  The series of processes described above can be performed by dedicated hardware or software. When a series of processing is performed by software, a program constituting the software is installed in an information processing apparatus such as a general-purpose computer or a microcomputer, and the information processing apparatus functions as the speech synthesizer 100.

  The above program can be recorded in advance in a hard disk drive (HDD) or ROM as a recording medium built in the computer.

  Alternatively, the program is not limited to a hard disk drive, but a removable recording medium such as a flexible disk, CD-ROM (Compact Disc Read Only Memory), MO (Magneto Optical) disk, DVD (Digital Versatile Disc), magnetic disk, and semiconductor memory. In addition, it can be stored (recorded) temporarily or permanently. Such a removable recording medium can be provided as so-called package software.

  The program is installed on the computer from the removable recording medium as described above, and is transferred from the download site to the computer wirelessly via a digital satellite broadcasting artificial satellite, or a LAN (Local Area Network) or the Internet. Such a program can be transferred to a computer via a network, and the computer can receive the program transferred in this way and install it on a built-in hard disk drive.

  Here, in this specification, the processing steps for describing a program for causing a computer to perform various processes do not necessarily have to be processed in time series in the order described in the flowchart, but in parallel or individually. This includes processing to be executed (for example, parallel processing or processing by an object).

  The program may be processed by one computer, or may be distributedly processed by a plurality of computers.

  As mentioned above, although preferred embodiment of this invention was described referring an accompanying drawing, this invention is not limited to this example. It is obvious for a person skilled in the art that various changes or modifications can be envisaged within the scope of the technical idea described in the claims, and these are naturally within the technical scope of the present invention. It is understood that it belongs.

  In the above embodiment, the case where the text analysis unit 101 and the prosody prediction unit 103 are separate bodies has been described as an example, but the present invention is not limited to such an example. For example, the present invention can be implemented even when the text analysis unit 101 and the prosody prediction unit 103 are configured integrally.

  The present invention is applicable to a speech synthesizer, a speech synthesis method, and a computer program capable of synthesizing speech for reading a text body.

1 is a block diagram showing a schematic configuration of a speech synthesizer according to a first embodiment. It is a flowchart which shows the outline of the speech synthesis method by the speech synthesizer concerning 1st Embodiment. It is explanatory drawing which shows the outline of the combination of the phoneme symbol concerning this Embodiment, prosodic prediction information, and the phoneme candidate of the whole text body. It is explanatory drawing which shows roughly the flow of the data in the speech synthesizer concerning this Embodiment. It is explanatory drawing which shows an example of the outline of the process which selects the phoneme concerning this Embodiment. It is a flowchart which shows the outline of the keyword priority phoneme selection process concerning 1st Embodiment. It is a flowchart which shows the outline of the keyword priority phoneme selection process concerning 1st Embodiment. It is a flowchart which shows the outline of the keyword priority phoneme selection process concerning 2nd Embodiment. It is a flowchart which shows the outline of the keyword priority phoneme selection process concerning 2nd Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Speech synthesizer 101 Text analysis part 103 Prosody prediction part 104 Phoneme selection part 107 Keyword priority phoneme selection part 109 Corpus 111 Phoneme connection part

Claims (8)

A prosodic prediction information generating unit that analyzes the text body and generates prosodic prediction information that predicts the feature amount of each symbol constituting the text body;
A scale of each symbol constituting the text body is obtained by mapping from a observable feature amount to a scale determined based on a perceptual experiment based on the prosodic prediction information and each symbol constituting the text body. A scale calculator;
A phoneme database storing one or more phonemes constituting the recorded speech;
A phoneme selection unit that selects a text phoneme from the phoneme database based on a scale obtained by the scale calculation unit for each symbol constituting the text body;
Each symbol constituting the keyword is mapped from the observable feature amount to a scale determined based on the perceptual experiment based on the prosodic prediction information and each symbol constituting the keyword in the text body. A keyword scale calculator for calculating the scale of
A keyword priority phoneme selection unit that selects a keyword phoneme from the phoneme database based on a scale of each symbol that constitutes the keyword obtained by the keyword scale calculation unit for each symbol that constitutes the keyword;
A phoneme replacement unit that replaces a phoneme corresponding to the keyword portion in the text phoneme selected by the phoneme selection unit with the keyword phoneme selected by the keyword priority phoneme selection unit;
A synthesis unit that connects the phonemes and generates synthesized speech;
A speech synthesizer characterized by comprising: When there are two or more keywords in the text body, the keyword priority phoneme selection unit assigns a priority to each of the keywords in the order in which the keywords appear from the top of the text body, 2. The speech synthesizer according to claim 1, wherein weights assigned to keywords are converted into priorities, and the keyword phonemes are selected in the order of priority. When there are two or more keywords in the text body, the keyword priority phoneme selection unit weights each keyword unit and, in the order in which the keywords appear, based on the weighted values of the keywords, The speech synthesizer according to claim 1, wherein a scale of each symbol constituting the symbol is obtained. The phoneme replacement unit sequentially replaces the phoneme corresponding to the keyword part in the text phoneme selected by the phoneme selection unit with the keyword phoneme when the keyword phoneme selection unit selects the keyword phoneme. The speech synthesizer according to claim 1, characterized in that: The phonological replacement unit determines whether or not to replace with the keyword phonology by comparing the metric obtained by the keyword priority phonological selection unit with the metric obtained by the phonological selection unit. 5. A speech synthesizer according to claim 1 or 4. The phonological replacement unit compares a scale value obtained by the keyword priority phonological selection unit and a difference value between the scale obtained by the phonological selection unit and a predetermined threshold value. Or the speech synthesizer according to claim 5. A prosodic prediction information generation step of analyzing the text body and generating prosodic prediction information in which the feature amount of each symbol constituting the text body is predicted;
A scale of each symbol constituting the text body is obtained by mapping from a observable feature amount to a scale determined based on a perceptual experiment based on the prosodic prediction information and each symbol constituting the text body. A scale calculation step;
For each symbol constituting the text body, based on the scale obtained in the scale calculation step, phoneme selection for selecting a text phoneme from a phoneme database storing one or more phonemes constituting the recorded speech Steps and;
Each symbol constituting the keyword is mapped from the observable feature amount to a scale determined based on the perceptual experiment based on the prosodic prediction information and each symbol constituting the keyword in the text body. A keyword scale calculating step for obtaining a scale of
A keyword priority phoneme selection step for selecting a keyword phoneme from the phoneme database based on the scale of each symbol constituting the keyword obtained in the keyword scale calculation step for each symbol constituting the keyword;
A phoneme replacement step of replacing a phoneme corresponding to the keyword part in the text phoneme selected in the phoneme selection step with the keyword phoneme selected in the keyword priority phoneme selection step;
A synthesized speech generation step of connecting the phonemes and generating a synthesized speech;
A speech synthesis method comprising: A prosodic prediction information generation step of analyzing the text body and generating prosodic prediction information in which the feature amount of each symbol constituting the text body is predicted;
A scale of each symbol constituting the text body is obtained by mapping from a observable feature amount to a scale determined based on a perceptual experiment based on the prosodic prediction information and each symbol constituting the text body. A scale calculation step;
For each symbol constituting the text body, based on the scale obtained in the scale calculation step, phoneme selection for selecting a text phoneme from a phoneme database storing one or more phonemes constituting the recorded speech Steps and;
Each symbol constituting the keyword is mapped from the observable feature amount to a scale determined based on the perceptual experiment based on the prosodic prediction information and each symbol constituting the keyword in the text body. A keyword scale calculating step for obtaining a scale of
A keyword priority phoneme selection step for selecting a keyword phoneme from the phoneme database based on the scale of each symbol constituting the keyword obtained in the keyword scale calculation step for each symbol constituting the keyword;
A phoneme replacement step of replacing a phoneme corresponding to the keyword part in the text phoneme selected in the phoneme selection step with the keyword phoneme selected in the keyword priority phoneme selection step;
A synthesized speech generation step of connecting the phonemes and generating a synthesized speech;
A computer program for causing a computer to execute a speech synthesis method including

Images