JP2710045B2 - Voice recognition method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition method for making a machine recognize a human voice. 2. Description of the Related Art In recent years, word-speech recognition devices have become more and more popular because their capabilities have become larger and their size and price have been reduced. However, it is not an easy-to-use device because there are various restrictions on use, such as restrictions on the number of vocabularies that can be used, restrictions on the use environment, and words that must not be used. In particular, in a device for an unspecified speaker, even those who do not know these restrictions on use use the device. For this reason,
It is common to use a word-speech recognition device for an unspecified speaker in combination with a voice response device to form a system of a question answering type. That is, by guiding the content of the user's utterance by the guide voice emitted from the voice response device,
Consider using the device without being aware of any restrictions on its use. For example, when a train ticket is reserved using a device that recognizes the ten digits and the twelve words of “high” and “no”, the following is the case. FIG. 4 is a functional block diagram (conventional example) of a currently used recognition and response system for unspecified speakers. When the task of the host computer 51 is started, the voice synthesis unit 52 outputs a guide voice. When the user utters the voice according to the instruction of the guide voice, the voice is recognized by the voice recognition unit 50, and the recognition result is transmitted to the host computer 51. In the host computer 51, a response sentence is created based on the intention of the user, and the speech
52 outputs a response voice. The user performs the task according to a procedure such as the next utterance. The speech recognition unit 50 performs a recognition process as follows. First, the acoustic analysis unit 53 performs filter analysis, LPC analysis, power calculation, and the like, and extracts parameters. The voice section detection unit 54 detects a voice section of the input voice using the result of the acoustic analysis. The similarity calculation unit 55 calculates the similarity between the parameter included in the voice section and the standard pattern of each voice stored in the standard pattern unit 56, and calculates a word corresponding to the standard pattern having the highest similarity. Is the recognition result. Problems to be Solved by the Invention As described above, if the speech recognition device is used in a question answering format, it can be used by anyone who has some understanding of the speech recognition device. However, those who are completely unaware of the limitations of speech recognition devices do not always respond in the expected way. For example, in the case of Example 1 above, the user may reply as underlined. Example 2: Equipment: How many tickets do I need? Please answer in numbers. User: Well, four. Equipment: Shinkansen ticket 4 User: Yes. The response of the user in Example 2 is an expression that is used naturally in conversation, but in speech recognition for word speech, such sentence speech cannot be recognized, so it is rejected or a recognition error occurs. Often. That is,
In the conventional word-speech recognition device, it is necessary to reply with a predetermined expression as in Example 1, and even a slight modification of the expression is not allowed. This is a very restrictive use,
This is a major cause of narrowing the field of use of the speech recognition device. An object of the present invention is to solve the problems of the conventional example and relax restrictions on the use of the word speech recognition device. In other words, the purpose of the present invention is to provide a means for correctly recognizing even if there are some extra parts other than the expression to be recognized in the expression of the input speech, thereby alleviating the restriction on the use of the device. And Means for Solving the Problems The present invention achieves the above object, and its technical means is to analyze an input signal including noise and unnecessary sound and convert it to an input parameter time series, and The similarity between the partial section of the sequence and the standard pattern of the recognition target word created in advance is continuously calculated while shifting the partial section by the unit section from the beginning to the end of the input parameter time series. One or more recognized word candidate sections that become larger are cut out, and the word name corresponding to the section is set as a recognized word candidate. On the other hand, the pitch frequency and its temporal change pattern are continuously obtained separately from the input signal, Determining a recognition word from the recognition word candidates by using the magnitude of the pitch frequency in or near the recognition word candidate section and its change pattern. The speech recognition method is characterized by the following. Action The present invention does not consider the entire speech section as one word, but rather from a sufficiently wide section including the recognition target word (including an extra word before and after the recognition target word and noise before and after the speech). Only the recognition target word is cut out and recognized using the value of the similarity by pattern matching, the magnitude of the pitch frequency, and the movement. That is, for a part of a sufficiently wide section including the recognition target word, the similarity to each of the standard patterns of the recognition target word is calculated, and then the similarity is calculated by shifting the unit section by unit. .., The movement of the similarity with each standard pattern is obtained for the entire area of the section.
Then, a section (or a plurality of sections) having a large similarity is obtained. Next, refer to the pitch frequency of the section with the highest similarity,
Only when the pitch frequency shows a value that is stably higher than around the section, the section is recognized as a section of the recognition target word. If the pitch frequency is not steadily higher than the surroundings, the section is rejected. In this way, by using the method of spotting and recognizing only the part of the word to be recognized, it is possible to obtain a correct and correct recognition result even for an input including extra words and noises, and a word speech recognition apparatus. Can be eased, and an easy-to-use voice recognition response system can be realized. In addition, this makes it possible for even a person unfamiliar with the device to use the device, thereby expanding the applications of the voice recognition device. Examples Hereinafter, examples of the present invention will be described. The present invention cuts out keywords from words uttered by pattern matching (word spotting),
This is a method of recognizing this by verifying it using a pitch frequency change pattern. The word spotting is a method of calculating the similarity while scanning each word standard pattern over the entire area of the input voice, and extracting a section having a high similarity and a word name at that time. However, since the word extracted as the maximum similarity is not always the correct answer, a certain criterion is set, and a plurality of candidates satisfying the criterion are extracted. And from among those candidates,
Using the pitch frequency change pattern and the similarity value, narrow down to one word. The pitch frequency is used to take advantage of the fact that humans speak higher frequencies and more clearly in important words than in others. Especially in conversation, the tendency is strong due to the experience of voice analysis. In a question answering sentence as in Example 2, since an important word (keyword) is a recognition target word, it is possible to determine the position where the recognition target word exists by using the pitch frequency. FIG. 1 is a functional block diagram of a recognition response system using a speech recognition method in one embodiment of the present invention based on such a concept. In the figure, since the functions of the host computer 51 and the speech synthesizer 52 are exactly the same as those of the conventional example, only the contents of the speech recognizer 6 will be described. The input speech is divided into analysis sections (frames) by the acoustic analysis unit 1, and is subjected to LPC analysis for each frame to extract LPC cepstrum coefficients. 8K sampling frequency
A Hamming window of Hz, a frame period of 10 msec, and an analysis window length of 20 nsec is used. The order of the LPC analysis is a 10th order, and the LPC cepstrum uses a fifth order (C _{1 to} C ₅ ) coefficient and a power term C ₀ . The pitch extracting unit 4 calculates a pitch frequency for each frame and accumulates the value for a certain period. There are various pitch extraction methods, but the simplest waveform correlation method is used. When the input signal is X _i , the correlation function μ _τ is _τ when μ _τ is the maximum Then the pitch frequency is Is required. The similarity calculation unit 2 sequentially compares the input parameter (LPC cepstrum coefficient) with the standard pattern of each word in the word standard pattern unit 3, and cuts out a portion having a high similarity as a word and stores it. The similarity calculation is performed not only in a place where a voice exists but also in a sufficiently wide section including a preceding and following noise section, and it is unnecessary to detect a voice section. Next, a method of spotting a word from noise or voice by similarity calculation will be described. First, a distance scale (statistical distance scale) used for pattern matching will be described. The input word speech length is linearly signaled into J frames of a fixed length,
Parameter vector per frame Then the input vector Is as follows. Here, t represents transposition. Where each Is a p-dimensional vector. As a standard pattern of the word ωn (n = 1, 2,..., N), the average vector is Covariance matrix What is necessary is just to make the word to be the recognition result. From Bayes' theorem P (ωn) of the first term on the right side can be regarded as a constant. Assuming a normal distribution, the second term is Denominator term Can be regarded as a constant if the input parameters are the same, but will not be a constant when compared with different inputs.
here, Is the average Covariance matrix Is assumed to follow a normal distribution of (3) By taking the logarithm of the equation, omitting the constant term, and setting this as Mn, That is, Expanding equation (6) as However, Equation (8) is a primary discriminant with a small amount of calculation. here,
Equation (8) is modified as follows. That is, Mn is the partial similarity for each frame Is obtained by adding J times and subtracting once. Next, a description will be given of a method of spotting and recognizing a voice using the above distance scale and a method of reducing the amount of calculation. In word spotting, a reference point i is set in a sufficiently long section that certainly includes a voice to be recognized, and various subsections are considered based on i. The degree may be obtained by equation (11), and a word having a high degree of similarity throughout all the partial sections may be set as the recognition result for the reference point i. And i is 1 to I
The same operation may be performed by advancing unit intervals in the range of. In the present embodiment, words whose similarity falls within the third place and the section thereof are obtained. If this similarity calculation is executed as it is, the amount of calculation will be enormous.However, by limiting the subsection length in consideration of the duration of the word and using the partial similarity dj commonly during the calculation, Calculation amount can be reduced. FIG. 2 is an explanatory view of the method. When matching the input with word n,
This shows how the similarity is calculated by linearly expanding and contracting the partial section length l (l ₁ <l <l ₂ ) to the standard pattern length J, and fixing the end at each frame. The similarity is calculated by equation (11) along a route starting at point T on QR and ending at P. Therefore, all the similarity calculations per frame are performed within ΔPQR. By the way, in equation (11) Since the component is the j-th frame component after the expansion and contraction of the section length l, there is a corresponding input frame i '. Therefore, using the input vector, dj can be expressed as follows. Here, rk (j) is a function for relating the word length k and the linear expansion / contraction of J. Therefore, each frame of the input (11) can be easily calculated by selecting and adding the partial similarity having the relationship of i ', if the partial similarity with is previously obtained. By the way, since ΔPQR moves to the right every frame, Is calculated in the memory corresponding to ΔPQS and is shifted for each frame. If all the necessary similarities are in the memory, the partial similarity Can be largely omitted, and the amount of calculation becomes very small. The judgment unit 5 applies the pitch frequency change pattern (temporally smoothed) extracted by the pitch extraction unit 4 to the sections (three in this embodiment) cut out by the similarity calculation unit 2. Then, the number of recognized words is reduced to one. That is, in the cut-out section, (1) if the highest part of the pitch frequency is included in all the sections, the word corresponding to the section is regarded as the recognition result. (2) If there is no section corresponding to (1), there is a section in which the movement of the pitch frequency is convex in the section, and if the pitch frequency is sufficiently high (the peak of the convex mountain is the second). , And the unit corresponding to the section is set as the recognition result. (3) When there are a plurality of sections corresponding to (1) or (2), the one having the higher similarity is given priority. If there is no corresponding section, it is treated as a reject. Next, the above description will be described in a concrete example with reference to FIG. In FIG. 3, (a) shows the contents of the utterance corresponding to the time, and uses "Uh, four cards" in Example 2. (B) and (c) show the temporal movement of the power and the pitch frequency, respectively. The pitch frequency forms a peak at the portion indicated by the thick line 31, and the pitch frequency has a maximum value over the entire region at the peak portion. (D) is a temporal movement of the similarity with respect to each standard pattern, and the thick line 32 to 34 is a section cut out by the similarity calculation unit 2 (according to the method described with reference to FIG. Since a portion having a large similarity is located at the end of the above, this is also expressed in FIG. 3). The above two conditions (1) and (2) are applied to each of these three sections. First, 34 is rejected because the pitch frequency is not the maximum and the movement is not convex. Although the pitches of 32 and 33 are both convex, since the pitch frequency is the maximum in the section of 33, the word “Yon” corresponding to 33 is set as the recognition result. In this way, a correct recognition result can be obtained from extra words and ambient noise. As described above, by using the method of this embodiment, a target word can be correctly spotted from noise or simple sentence speech. Since there is no need to detect a voice section in advance, the processing is simple. Moreover, since the amount of calculation required for the similarity calculation is small, hardware implementation is very easy. Advantageous Effects of the Invention As described above, the present invention cuts out the recognition word candidates and their sections from the input speech by the continuous linear expansion / contraction matching method, obtains the pitch frequency and its change pattern, and determines the pitch frequency within the recognition word candidate section. When the frequency is high enough and the change pattern is convex, it identifies the recognition word, and the conventional device must be used in a very quiet environment, and the user can speak other words than necessary. This has the advantage that the restrictions on use can be relaxed and the use of the word speech recognition device can be expanded and spread.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a functional block diagram of a recognition response system which embodies a speech recognition method according to an embodiment of the present invention, and FIG. FIGS. 3 (a) to 3 (d) are conceptual diagrams illustrating a specific example of the recognition method according to the present embodiment, and FIG. 4 is a diagram illustrating a conventional speech recognition method. It is a block diagram of a recognition response system. 1 ... Acoustic analysis unit, 2 ... Similarity calculation unit, 3 ... Word standard pattern unit, 4 ... Pitch extraction unit, 5 ... Judgment unit.

Claims (1)

(57) [Claims] The input signal including noise and unnecessary sound is analyzed and converted into an input parameter time series, and the similarity calculation between the partial section of the input parameter time series and the standard pattern of the recognition target word created in advance is performed. Is continuously performed while shifting the unit section from the beginning to the end of the input parameter time series, and one or more recognized word candidate sections having a high similarity are cut out, and a word name corresponding to the section is set as a recognized word candidate. On the other hand, separately, the pitch frequency and its temporal change pattern are continuously obtained from the input signal,
A speech recognition method, wherein a recognition word is determined from the recognition word candidates by using a magnitude of a pitch frequency in or near the recognition word candidate section or a change pattern thereof. 2. As a method of calculating the similarity between the input parameter time series and the standard pattern, the method is characterized in that the time length of the input parameter is linearly expanded and contracted to the time length of the standard pattern, and is calculated using a statistical distance scale that is posterior-probable. Claims 1
The speech recognition method described in the section. 3. As a method of determining a recognition word from among the recognition word candidates, in the recognition word candidate section, when the pitch frequency takes a maximum or equivalent value stably in the entire input signal section, or in the recognition word candidate section, , When the pitch frequency change pattern becomes convex,
2. The speech recognition method according to claim 1, wherein a word corresponding to the recognition word candidate section is set as a recognition word.