JPH04347898A - Voice recognizing method - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method suitable for recognizing speech uttered in a place with a lot of ambient noise.

0002

2. Description of the Related Art As a means of transmitting information, voice is generally superior to manually operated typewriters, push buttons, etc. in terms of transmission speed and operability. When using voice as a means of transmitting information, if the recipient is a machine, a voice recognition device is required to correctly recognize the content of the uttered voice.

However, speech recognition devices are used not only in quiet rooms, but also in noisy environmental conditions that degrade recognition performance. Therefore, as a method to improve recognition performance, it is common to perform recognition processing after eliminating the influence of noise as much as possible to make the sound clean. Techniques have been proposed, such as configuring a two-input system by arranging directional microphones with their directivity axes in opposite directions, and removing noise using the subtraction method.

[0004] However, even if the above-mentioned method is adopted, it is difficult to remove only the noise with necessary and sufficient characteristics from voices collected in a place with a high noise level. It is extremely difficult to obtain. On the other hand, when recognizing speech with superimposed noise,
It is known that recognition performance is improved by performing recognition using a standard pattern created using speech with some noise added, compared to using speech without noise as the standard pattern. Therefore, we first conducted an experiment to confirm the effect on recognition when noise-added speech is used as a standard pattern, which is a prerequisite for realizing the improvement in recognition performance of this invention. Figure 3 shows the sound that was uttered by repeating a word set consisting of 100 city names twice when the noise level was changed from high to medium to low. A standard speech pattern was created using one word set and the other A standard speech pattern is created using the voice of the word set uttered when the noise level is low and when the word set is recognized by DP matching (solid line in the figure) using the voice as a test voice.
This figure shows a comparison between the case where the speech of the word set uttered under each noise condition is recognized as the test speech (dotted line in the figure).

[0005] From now on, when a standard speech pattern created using one of two speeches uttered under the same noise level conditions is used to recognize the other utterance as a test speech, the recognition rate obtained is While a value close to 100% can be obtained even when the noise level is high, standard speech patterns are created using sounds uttered when the noise level is low, and sounds uttered under medium and high noise levels. It can be seen that the recognition rate decreases as the noise level increases, and is reduced to about 30% for voices uttered under conditions of high noise level.

[0006] The results of this recognition experiment show that when recognizing speech with superimposed noise, using a standard speech pattern created using speech uttered under the same noise conditions is effective in improving the recognition rate. It was confirmed that this was the case. However, in order to realize this, it is necessary to record speech for creating a standard pattern each time the recognition device is used in an actual environment in which the speech recognition device is used. This is because when the environmental conditions in the room you use change, or when the type and level of noise changes even within the same room, it is necessary to utter a sound to create a standard pattern each time. Using a recognition device under conditions that satisfy these conditions requires a great deal of effort and time, and cannot be said to be a realistic method.

[0007]

[Problems to be Solved by the Invention] In view of the above points, the present invention provides a method for creating a standard speech pattern each time a recognition device is used when recognizing speech uttered under various noisy environment conditions. The purpose of the present invention is to provide a speech recognition method that can obtain practically sufficient recognition performance without the need for utterance.

[0008]

[Means for Solving the Problems] According to the present invention, the speech level for creating a standard speech pattern is changed in several stages in an anechoic room or a soundproof room without noise or reverberation. A word set sequence data obtained by cutting out only the voice part of a clean voice uttered is stored in advance in a buffer in the voice recognition device as a reference voice.

Noise data collected immediately before the test voice uttered as a word to be recognized in an actual use environment and its time average level Pn; and a test voice to be recognized with superimposed noise and its time average level Pt. and the ratio Pt /Pn of these, and the noise data collected in the actual usage environment when the test voice was uttered was added to the reference voice that was uttered with the level changed in several stages in advance, and the Pt /Pn was calculated. The level of the reference sound is adjusted and added so that it is equal to the ratio of Pn, and this addition is performed for all of the reference sounds uttered at each level, and the result is created for all of the additions of each reference sound and the noise data. A standard template for pattern matching is created from the standard speech generated, and test speech uttered in noise is recognized.

0010

EXAMPLES Examples of the present invention will be described in detail below. FIG. 1 shows an embodiment of this invention. The test voice data collection unit 11 collects voice data of the object to be recognized, which is uttered in a noisy actual usage environment and has noise superimposed thereon. In this test audio data, noise data on which no audio is superimposed appears alternately with audio data in time series. From this test audio data, each section of speech and noise is identified and detected by the speech section detection section 12. Speech sections can be detected by methods that focus on the difference in sound power level between speech sections and noise sections.For example, the threshold value is detected using clues such as the length of time that the threshold continues at a level higher than a certain level setting value. It is done. Test speech Si to be recognized detected as speech by the speech section detection unit 12, noise data Ni detected as noise, and these test speech Si
The acoustic power level Pt of the noise data Ni, the acoustic power level Pn of the noise data Ni, and their ratio Pt/Pn are stored in the test data storage section 13.

On the other hand, the clean voice data collection unit 14 for creating a standard voice pattern sets the voice level to several levels (for example,
The uttered voice is recorded in three stages (loud, medium, and soft). The necessity of recording while changing the vocalization level is due to the following reason. In other words, when a standard pattern is created and recognized using the voice uttered when the noise level is low as shown in Figure 3, the recognition performance for the voice when the noise level is high is degraded (Figure 3). (Characteristics indicated by the dotted line in 3) The reasons include the difference in spectral characteristics from clean voice due to noise being superimposed on the voice, as well as the difference in spectral characteristics when vocalizing in a noisy environment, depending on the noise level. This is because it is thought that the factor of vocal deformation (Lombard effect) that occurs due to different levels has a large influence.Therefore, assuming the case where the noise level changes in several stages, the voice with vocal deformation is the standard pattern. Required when creating.

[0012] Vocalization transformation is performed to create a standard speech pattern by applying the amount of change in physical characteristic components due to vocalization modification, such as the shift amount of formant frequency, to clean speech to normally uttered speech. In principle, it is possible to obtain a voice, and if the characteristics of the voice obtained by this change in several stages and have sufficient similarity to the voice uttered, it can be used to create a standard pattern. This method has the advantage that the voice only needs to be uttered once at a normal voice level, making it a very effective method.

[0013] The voice section detecting section 15 detects the voice section of the clean voice data with vocal deformation obtained using any of these methods, and the voice data that has been detected is converted into the reference voice Csij. and the word utterance time length is set as Tsij, and both are stored in the reference speech storage unit 16 for each difference in utterance level. Csij and Tsij
The subscript s indicates the reference voice for creating the standard pattern, and i indicates the difference in vocalization level (i = 1 to m, m≧2)
, and j indicates the recognized word number (j=1 to n), respectively. Such reference voices are stored in advance.

[0014] In recognition, each reference voice Csij is
The noise data Ni stored in the test data storage section 13 is superimposed in the addition section 18 via the level control section 17. At this time, there are two cases in which the time length T1 of the noise data Ni is longer and shorter than the time length Tsij of the clean audio data, and the noise superimposition method is performed as follows depending on each case.

This situation will be explained using FIG. 2. Figure 2A
shows the level change characteristics of the test voice data over time, where the time length of the noise data Ni is T1, its time average power level is Pn, followed by the test voice Si to be recognized, The utterance time length is T2, and the time average power level of the test voice including the noise is Pt. FIGS. 2B and 2C show that the above noise data (time length T1) Ni is added to the reference voice Csij.
This figure shows how the adding unit 18 adds the numbers. Psij is the average power level in the time length Tsij of the voice section when the above-mentioned noise having the characteristics of time average level Pn and time length T1 is added to the reference voice Csij.

[0016] When noise is superimposed here, ■T1>T
sij ■There are two conditions: T1≦Tsij. In the case of ■, as shown in FIG. 2B, the beginning of the noise Ni is set so that the time interval Tsij of the reference sound (in this figure, the cases of j=1 and 2 are shown) falls within the time T1 of the noise. The noise Ni may be superimposed by arbitrarily setting the time length Tp from the beginning to the beginning of the reference speech Csij to a value that allows the speech section to be reliably cut out when the speech section is detected later. At this time, Psij/Pn
The calculation unit 19 calculates the value of Pt /Pn stored in the test data storage unit 13, and the level comparison unit 21 compares this value with the value of Pt /Pn stored in the test data storage unit 13 so that the two become equal, that is, Psij = Pt. The power level Psij of the noise superimposed reference sound is adjusted by the level control unit 17 as follows.

On the other hand, in the case of ■, as shown in FIG. 2C, there is a section containing only noise that includes the time length Tsij of the reference voice (in this figure, the case where j=1 is shown) and before and after it. It is necessary to superimpose the voice and noise so that they exist. Therefore, for example, if the time length T1 of the noise is only about 1/3 of the time length Tsil of the reference sound, the same noise data having the time length T1 may be repeated 4 to 5 times and superimposed on the reference sound signal. becomes. At this time, the level Psij is controlled so that the value of Psij/Pn becomes equal to the value of Pt /Pn stored in the test data storage section 13, that is, Psij=Pt, as in the case of (2). Adjustment is made in section 17.

The standard audio data with the reference audio level adjusted in this way and the noise data Ni superimposed thereon is stored in the buffer 22. The standard voice data is input to the voice section detecting section 24 after the switch 23, which had been open until then, is closed. The algorithm and threshold value of the voice section detecting section 24 are set to be the same as the conditions of the voice section detecting section 12 for the previous test voice data, and the voice section is detected. By doing this, even if some errors occur in detecting speech sections, it is expected that the errors will be the same for both the test speech and the standard speech, and as a result, the present invention has a feature that can contribute to improving recognition performance.

The speech signal whose speech section has been detected by the speech section detection section 24 is subjected to speech spectrum analysis and creation of a standard pattern at the standard pattern creation section 25. In this way, a standard pattern is created by superimposing noise on the reference speech Csij for all recognized words at all utterance levels. The pattern matching unit 2 compares all these standard patterns with the test audio pattern created by the test pattern creating unit 26.
It will be held at 7. Here, a distance calculation is performed to determine the similarity between words, and the standard speech patterns are output as the recognition results of the test speech in the order of the highest similarity to the test speech, and the results of the first rank are used as necessary. In addition, the recognition result is outputted to the recognition result output unit 28 along with the ranking below, and a series of operations as a recognition device are performed.

[0020] Note that the test speech to be recognized described here is word speech, but due to the structure and method of the present invention, the speech to be recognized is not limited to word speech. do not have.

[0021]

[Effects of the Invention] As explained above, by using the speech recognition method according to the present invention, highly accurate recognition performance is ensured by performing recognition processing that corresponds to changes in environmental conditions when using a recognition device. You can. In other words, if the distance between the speaker and the microphone is close during actual use, the vocalization level will be small even if the noise level is somewhat high.Therefore, the amount of vocalization distortion will be small, but the S/N will be lower; - When the distance between the microphones is large, the speaker often unconsciously tends to speak loudly even when the noise level is low, as well as when the noise level is high. In order to deal with this, multiple sets of reference sounds with different amounts of vocalization distortion are prepared in advance, and the noise occurring in the actual usage environment is superimposed on the reference sound mentioned above. S of
/N value is set equal to the S/N value of the test voice, and the same algorithm and threshold as the test voice are used to detect the voice section for creating a standard pattern from the reference voice. As a result, the test voice and the standard voice can have almost the same characteristics as those uttered under the same environmental conditions. As a result, the present invention has the feature that it is possible to realize a speech recognition device that ensures a practically sufficient recognition accuracy by imposing only the minimum necessary burden on the user.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an example of a functional configuration of a speech recognition device to which the present invention is applied.

FIG. 2 is a diagram for explaining a method of superimposing noise data on clean audio data.

FIG. 3 is a diagram showing the relationship between noise level and recognition rate.

Claims (2)

[Claims] [Claim 1] Only speech segments are detected from clean speech data in which the word set of the vocabulary to be recognized is uttered at multiple levels of utterance under environmental conditions with few echoes and reverberations and no noise. The reference sound and its time average level are memorized, and the sound section and the noise section are detected from the input test sound data, and the test sound, noise data, and the time average level Pt of the test sound are determined.
, the time average level Pn of the noise data, and the ratio Pt /Pn of these, and control the level of the reference sound so that it becomes equal to the ratio Pt /Pn for all of the reference sounds, and compare it with the noise data. Create noise-superimposed speech by adding them, detect speech sections from these noise-superimposed speeches, create standard speech patterns for each word from these speech sections for each of the above pronunciation levels, and compare this standard speech pattern with the above test speech. A speech recognition method that recognizes test speech by performing pattern matching. 2. The audio according to claim 1, wherein the detection of the audio interval from the test audio data and the detection of the audio interval from the noise superimposed audio are performed using the same algorithm and threshold value. Recognition method.