JP2003241788A - Voice recognition device and voice recognition system - Google Patents

Abstract

An object of the present invention is to provide a speech recognition apparatus and a speech recognition system capable of accurately recognizing a user's speech input under a plurality of different noise environments with a low processing load. It is to be. An object of the present invention is to provide a speech recognition apparatus for recognizing a user's voice input under a plurality of different noise environments, wherein a feature of the user's utterance is made to correspond to each property of the plurality of different noises. Noise adaptive acoustic model holding means for holding a plurality of modeled acoustic models, and when the start of use of the speech recognition device is detected, the property of the noise is detected, and the property of the noise is detected. Voice recognition processing means for acquiring a voice model from the noise-adaptive voice model holding means, and performing voice recognition by comparing the voice model with the input user voice. Will be resolved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to voice recognition in a noisy environment, and more particularly, to a voice recognition device and a voice recognition system capable of performing voice recognition based on a noise state in a portable environment. .

[0002]

2. Description of the Related Art Generally, in a device which is expected to be used in a noisy environment such as a mobile device for mobile communication, it is essential to have a noise resistance technique for correctly recognizing a voice generated in a noise. In the field of speech recognition, many noise resistant techniques have been proposed, and typical techniques thereof include noise cancellation techniques and noise adaptation techniques. Noise cancellation technology is defined as the inverse Fourier transform of the logarithm of the noise spectrum or the cepstrum (the short-time amplitude spectrum of the speech waveform. This is called lifter operation, and the spectrum envelope component that represents the phonological property and the fundamental frequency that represents the pitch. There is a method of removing noise by utilizing the above bias. On the other hand, there is a noise adaptation technique in which noise itself is incorporated into an acoustic model and the acoustic model is modified to improve the recognition rate of speech including noise.

In addition to the above, there has been proposed a noise reducing method in which the directivity of the microphone is strengthened to prevent background noise from being mixed into the input voice.

As described above, there is a noise reduction technique as a technique for correctly recognizing a voice uttered by a user under noise. In an environment where a mobile device for mobile communication is used (hereinafter referred to as a mobile environment), since the mobile device is used in various places such as on the side of a road or at a platform of a station, a background superimposed on voice is used. Noise varies greatly depending on the place of use. Therefore, like the noise law method, in order to correct the acoustic model from the background noise immediately before the utterance using one noise adaptive model, the tracking performance of the correction in the mobile environment where the magnitude of the background noise changes momentarily. Is limited. However, if a method is adopted in which a large number of acoustic models adapted to a large number of individual background noises are retained, and which of these is determined and the acoustic model considered to be optimal is adopted, practical use in a mobile environment is possible. However, on the other hand, since a large amount of memory and processing capacity are required, the processing load is increased in a communication terminal, for example, a mobile device such as a mobile phone, which is not feasible.

Further, in cutting out a voice section before starting voice recognition, it is generally widely used to judge a voice start point by observing a change in a loudness (voice power) of a voice, but background noise is large. In a mobile environment, there is a problem that it is difficult to distinguish between voice and noise and it does not operate normally.
On the other hand, there is a solution using a microphone that eliminates background noise, but in this case, noise elimination causes voice distortion, and as a result, the recognition rate of voice is reduced. There is also a method of avoiding background noise by using a microphone with high directivity, but in this case, it is necessary to match the directivity with the direction of the user's mouth, and if the user is not aware of this, the accuracy of High speech recognition rates are not guaranteed. This means that the convenience for the user is reduced, and there is a problem in terms of practicality.

The present invention has been made in view of the above problems, and its object is to accurately recognize the voice recognition of a user input under a plurality of different noise environments with a low processing load. A voice recognition device and a voice recognition system that can be performed.

[0006]

In order to solve the above-mentioned first problem, the present invention, as set forth in claim 1, recognizes a user's voice input under a plurality of different noise environments. In a voice recognition device, noise adaptive acoustic model holding means for holding a plurality of acoustic models in which the characteristics of the user's utterance are modeled corresponding to the properties of each of the plurality of different noises, and the use of the voice recognition device is started. Is detected, the property of the noise is detected, an acoustic model corresponding to the detected property of the noise is acquired from the noise adaptive acoustic model holding means, and the acoustic model and the input user are acquired. And a voice recognition processing means for performing voice recognition by collating the voices.

According to such a speech recognition apparatus, a plurality of acoustic models corresponding to the nature of noise (= background noise), for example, the volume of background noise are held in advance, and the use of the speech recognition apparatus by the user is detected. The acoustic model corresponding to the volume of the background noise detected at the time of being selected is selected from the retained acoustic models. Since the recognition of the voice input by the user is performed by adopting the acoustic model selected as described above, even if there is a change in the noise environment due to the movement of the user, the use of the voice recognition device by the user Previously, speech recognition is performed using an acoustic model that matches the noise environment at that location. That is, it is possible to obtain good recognition performance even when the voice recognition device is used in a mobile environment. Further, as the generation mode of the acoustic model, acoustic models associated with the noise volume may be prepared, or one original acoustic model may be adaptively changed according to the noise volume. In the latter case, a control circuit or the like for changing one acoustic model according to the noise volume without delay may be provided. Further, here, the example in which the acoustic model is generated in accordance with the noise volume is shown, but the present invention is not limited to this, and may be generated in accordance with the type of noise.

From the viewpoint that it is possible to obtain a high speech recognition rate while minimizing the processing amount by preparing two noise models, the present invention is as described in claim 2. In the voice recognition device, the acoustic model held in the noise adaptive acoustic model holding means is generated in correspondence with the level of noise volume representing the magnitude of the noise.

According to such a voice recognition device, an acoustic model adapted to a low noise volume (herein, abbreviated as a low noise acoustic model) and an acoustic model adapted to a high noise volume (here, a high noise volume). It is only necessary to prepare two models (abbreviated as acoustic model), so that voice recognition can be accurately performed with a low processing amount even in a noisy environment in which the noise volume greatly changes like a portable environment. . The determination of the start of use of the voice recognition device by the user is, for example, as described in claim 3, in the voice recognition device, the use start detection criterion of the voice recognition device is such that the user recognizes the voice recognition device. And when the voice recognition device is operated by the user, and when at least one of when the voice recognition device contacts or approaches a part of the user is detected.

It is said that by performing noise adaptive correction of either the low noise acoustic model or the high noise acoustic model based on the ambient noise situation before the start of speech, it becomes possible to recognize speech by a more realistic acoustic model. From the point of view, the present invention provides the speech recognition device as described in claim 4,
If the sound recognition unit measures the noise and the volume of the noise is small, the speech recognition unit acquires an acoustic model associated with the small noise amount from the noise adaptive acoustic model holding unit to obtain background noise immediately before utterance. If the volume of the noise is large, an acoustic model associated with the large noise amount is acquired from the noise adaptive acoustic model holding unit and corrected with the sound immediately before the utterance. And a second noise adaptive correction means for performing the above.

According to such a voice recognition device, in a quiet environment where background noise is relatively small, such as an indoor environment,
A low-noise acoustic model is adapted using background noise before the start of speech, and a high-noise acoustic model is applied based on the surrounding sound in a crowded environment. Speech recognition by the model becomes possible.

It is said that the recognition accuracy of the voice recognition can be further improved by performing the voice recognition using the analysis result of the voice analyzed according to the volume of the background noise and the acoustic model obtained as described above. From the point of view, according to the present invention, as described in claim 5, in the voice recognition device, the voice recognition means analyzes the voice input when the measured noise volume is low. Analytical means,
A second voice analysis means for analyzing a voice input when the measured noise volume is large, and the input voice analyzed by the first voice analysis means is subjected to the first noise adaptive correction. Recognize using the acoustic model corrected by means,
It is characterized in that the input voice analyzed by the second voice analysis means is recognized by using the acoustic model corrected by the second noise adaptive correction means.

According to the present invention, as described in claim 6, in the voice recognition device, the voice recognition means performs the measurement, from the viewpoint that the voice analysis according to the volume of the background noise can be reasonably performed. It is characterized by further comprising utterance start determination switching means for switching between the first speech analysis means used for analyzing the input speech and the second speech analysis means according to the magnitude of the noise.

According to such a voice recognition device, the voice analysis means used for determining the utterance start can be switched according to the measured noise volume, so that the utterance start point can be reasonably extracted.

From the viewpoint that voice recognition in the utterance section can be reasonably performed when the background noise is low, the present invention provides the voice recognition device according to the first aspect, as described in claim 7. The voice analysis means of has a voice power detection means for determining the start of utterance by using the power information of the input voice, and outputs the voice of the detected utterance section by the voice recognition means for recognition. There is.

According to such a voice recognition device, in an environment where the background noise is small, it is possible to adaptively reduce the processing amount by performing the utterance start detection at the volume (power) which is a simple process. It is possible to reduce power consumption.

From the viewpoint that the utterance start point can be accurately extracted even when the background noise is high, the present invention provides the speech recognition device according to the second aspect, as described in claim 8. The voice analysis means of 1 divides the input voice into two, and applies a cepstrum analysis to the voice storage means that stores one voice together with time information and the other voice, and the start point of the pitch component obtained by the analysis is the utterance start point. Pitch extraction means for determining as a candidate, the point before the obtained utterance start point candidate is regarded as the utterance start point, and the speech in the utterance section from the determined utterance start point is output by the voice recognition means. It is characterized by making them recognize.

According to such a voice recognition device, cepstrum analysis is performed in an environment with a large background noise,
Since the utterance starting point is estimated after extracting the characteristics of the voice and the pitch component, the utterance starting point can be accurately extracted even in a noisy environment such as an underground shopping area or a busy street.

Further, according to the present invention, as described in claim 9, in the voice recognition device, the voice recognition processing means calculates a recognition score indicating a recognition accuracy as a recognition result, and the recognition score is calculated. If the recognition score exceeds a predetermined threshold, the recognition result is considered to be highly likely to be correct, and a voice feature recognition result storage unit that stores the recognition result and a voice feature amount obtained by the recognition, A speaker adaptation adaptation unit for learning a speaker-specific voice feature using the recognition result stored by the voice feature recognition result storage unit and the voice feature amount obtained by the recognition. There is.

According to such a voice recognition device, for example, when the background noise is small, the recognition score is calculated as the voice recognition result for the uttered voice. This recognition score is expressed as a score of likelihood of recognition, and the higher the score (compared to a predetermined threshold), the more accurate the recognition at the time of voice recognition is. According to the present invention, only when a high recognition score is obtained, the recognition result and the voice feature amount obtained by the recognition are stored as data necessary for speaker application, so that the recognition target word is not limited. Sufficient recognition performance can be obtained.

From the viewpoint that the speaker adaptation can be performed at a higher speed by performing the speaker adaptation only when the voice recognition device of the present invention is connected to the external power source.
According to the present invention, as described in claim 10, in the voice recognition device, the speaker adaptation means stores the voice feature recognition result storage when the voice recognition device is supplied by an external power supply means. An external power supply speaker adaptation means for performing speaker adaptation using the recognition result stored by the means and the voice feature amount obtained by the recognition is provided.

According to such a voice recognition device, since the speaker adaptation processing is executed only when the speaker itself is supplied with power from the outside, as compared with the case where the speaker adaptation processing is driven by the battery. Not affected by battery life.
Therefore, the clock speed of the control unit can be increased,
The speaker adaptation process can be executed at high speed. further,
If the user makes the connection between the voice recognition device and the external power source at bedtime, the speaker adaptation process is completed while the user is sleeping, so that the convenience for the user can be improved.

The present invention is set forth in claim 11 from the viewpoint that it is possible to further reduce the processing load on the speech recognition apparatus by externally performing speaker adaptation processing that requires a large amount of processing. As described above, in the voice recognition system including the communication terminal capable of wireless or wired voice communication and the server device, the communication terminal is the voice recognition device according to any one of claims 1 to 10, and the voice recognition device. When the amount of data representing the voice recognition result about the input voice obtained by the above and the voice feature amount is accumulated for the application of the speaker, the data batch transmission means for transmitting the data to the server device is provided. The server device includes data receiving means for receiving the data transmitted from the communication terminal,
A data batch processing type speaker adaptation processing means for executing speaker adaptation based on the data, and a parameter to be changed in the communication terminal by the data batch processing type speaker adaptation processing means, and the communication terminal And a data batch processing type change parameter transmitting means for transmitting the data.

According to such a voice recognition system, when the corresponding data necessary for speaker adaptation is accumulated on the communication terminal side which is the client side, the data is sent to the external server side. The server device executes the speaker adaptation process based on the data received from the communication terminal, and returns the information of the parameter to be changed in the communication terminal. That is, the speaker adaptation calculation is executed on the external server side, and the result is fed back to the communication terminal on the client side.
It is possible to reduce the computer resources and memory resources required for the communication terminal. Further, even while the speaker is being adapted, the communication terminal can be used without being overloaded. Further, the communication mode between the communication terminal and the server device may be a wireless communication mode such as a wireless LAN, or a wired communication mode such as a USB. So
Providers will be able to implement voice recognition systems in a variety of communication formats. On the other hand, the user side can select a system suitable for each from various communication modes, which improves convenience.

From the same viewpoint as above, according to the present invention, as defined in claim 12, in the voice recognition system, the communication terminal performs voice recognition relating to an input voice obtained by the voice recognition device. A data receiving and storing unit that includes a data sequential transmitting unit that sequentially transmits a result and data representing a voice feature amount to the server device, and the server device receives and accumulates the data transmitted from the communication terminal. And a data-sequential accumulation processing type speaker adaptation processing means for executing speaker adaptation based on the data when the amount of the accumulated data reaches the amount necessary for the speaker adaptation, and the data-sequentially. Data for calculating a parameter to be changed by the communication terminal by the storage processing type speaker adaptation processing means and transmitting it to the communication terminal-sequential storage processing type change parameter transmitting means; It is characterized by comprising.

Further, from the same viewpoint as above, the present invention provides
As described in claim 13, in the voice recognition system, the communication terminal transmits predetermined data by the data batch transmission means or the data sequential transmission means,
It is characterized in that it is performed to the server device via a communication network connected to the server device.

According to such a voice recognition system, the speaker adaptation calculation is executed on the server side connected to a communication network such as a public network, a mobile network or an IP network, and the result is transmitted to a communication terminal on the client side. Since it is fed back, the shared resources of the communication terminal can be effectively used. Also, a provider such as a telecommunications carrier can realize a voice recognition system via a communication network.

If the communication network is a mobile network,
As described in claim 14, the communication terminal is characterized in that, in the voice recognition system, the communication terminal is a mobile communication terminal.

[0029]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of a voice recognition device according to this embodiment. This voice recognition device includes a use start / end determination unit 11, a voice input unit 12, a noise volume determination unit 13, an utterance start determination switching unit 14, a filter 15,
The voice storage unit 16, the pitch extraction unit 17, the voice start end detection unit 18, the noise adaptive correction unit 19, and the acoustic model storage unit 2
0, a voice recognition processing unit 21, a voice feature / recognition result storage unit 22, and a speaker adaptation unit 23.

In the above configuration, the acoustic model storage unit 20
Includes an acoustic model A that is noise-adapted with a small amount of background noise, such as a quiet indoor environment (= noise environment A), and the surrounding noise is large, such as in an underground city with many pedestrians or in a downtown area. Two types of acoustic models, that is, the acoustic model B that is noise-adapted with the volume of the background noise under the environment (= noise environment B) are held in advance. Below, an explanation will be given of the outline of the operation in the above-mentioned configuration on the assumption of this.

First, the use start / end determining section 11 detects the start of use of the voice recognition device by the user by the following means.

Grip Detection by Touch Sensor In the case of a folding type voice recognition device, an infrared sensor detects that the user's ear has touched the speaker part of the flip recognition voice recognition device. When the use start of the user is detected, the use start end determination unit 11 outputs a signal (= start instruction signal) indicating the use start by the user to the voice input unit 1.
Output to 2. Upon receiving this notification, the voice input unit 12 outputs the noise input before the start of speech to the noise volume determination unit 13. The noise volume determination unit 13 measures the volume of noise (hereinafter abbreviated as background noise) output from the voice input unit 12,
It is determined whether the measured volume of the background noise corresponds to the noise environment A or B, and the result is output to the voice recognition processing unit 21. The result determined by the noise volume determination unit 13 is also sent to the noise adaptive correction unit 19. The noise adaptive correction unit 19 holds two noise models based on the noise environments A and B, and corrects based on the determination result from the noise volume determination unit 13. The correction result of the noise model (the correction result, the corrected noise model itself, or the correction information only) may be output to the voice recognition processing unit 21.

The input voice input to the voice input unit 12 is
It is output to the voice recognition processing unit 21 through different voice analysis processing depending on the size of background noise. In this voice analysis process, for example, when the background noise measured by the noise volume measurement unit 13 is small, the voice start edge detection unit 18 extracts the utterance section and outputs the utterance section to the voice recognition processing unit 21. On the other hand, when the background noise is large, the input speech is filtered by the filter 15 and the pitch extraction unit 1.
The cepstral analysis is applied via 7 to extract the utterance section, the data of the section voice is read from the voice storage section 16 and output to the voice recognition processing section 21.

The speech recognition processing section 21 performs speech recognition of the speech section of the input speech, which is divided and extracted according to the magnitude of the background noise as described above, with the noise model output from the noise adaptive correction section 19, The recognition result is obtained by using the acoustic model stored in the acoustic model storage unit 20. After the voice recognition is performed as described above, the voice recognition processing unit 21 collects the teacher data necessary for speaker adaptation only when the background noise is small, and the teacher data obtained by the collection. Is output to and stored in the voice feature / recognition result storage unit 22. The speaker adaptation unit 23 adjusts and updates the acoustic model stored in the acoustic model storage unit 20 based on the teacher data stored in the voice feature / recognition result storage unit 22 so as to match the user's feature. That is, the voice recognition processing unit 2
In No. 1, since voice recognition can be performed using the acoustic model updated by the information of the noise environment immediately before the user speaks, a high recognition rate can be obtained.

A specific processing example of the speech recognition apparatus configured as described above will be described with reference to FIG. Figure 2
3 is a flowchart showing a processing procedure of a voice recognition process in the voice recognition device of the present invention.

As described above, in the acoustic model storage unit 20 in the voice recognition device, two types of noise models are prepared according to the volume of background noise, and the noise-adapted acoustic model is preliminarily used as the noise model. Retained. For example, two acoustic models in which noise adaptation is performed based on background noise of small background noise (hereinafter abbreviated as noise model A) and relatively large background noise having a wide band (hereinafter abbreviated as noise model B) Is held in advance. The noise model A is a background noise environment of a quiet office or a library, and the noise model B is a noise model modeled by background noise in an environment of an underground mall / shopping district with a lot of traffic.

First, in the use start / end detection section 11,
When the use start by the user is detected (S1) by any of the above-described four methods (above), the noise volume determination unit 13 measures the volume of the background noise (S2). The noise volume determination unit 13 compares the volume of the measured background noise with a predetermined threshold value (S3), and if the background noise is low (low in S3), the noise model A (S4), and if the background noise is high ( The noise model B (high in S3) (S8) is adopted and output to the noise adaptive correction unit 19. First, the processing flow when the noise model A is obtained by the determination by the noise volume determination unit 13 will be described.

(Processing Flow when Noise Model A (Low Noise Model) is Adopted) The noise adaptive correction unit 19 which has obtained the noise model A by the judgment by the noise volume judgment unit 13 is immediately before the utterance by the user (before the voice input is started). Adaptive correction (S5) of the noise model is performed based on the background noise of. In addition, the noise volume determination unit 13
When the noise volume determination is completed, the switch in the speech start determination switching unit 14 is switched according to the determination result of the noise volume determination unit 13. Here, since the noise volume determination unit 13 determines that the noise model A is adopted, the switch in the speech start determination switching unit 14 is switched to the voice start end detection unit 18 side. In the voice start point detection unit 18, the input voice from the voice input unit 12 is input, the start end of the voice uttered by the user is detected based on the volume of the sound (power of the sound) (S6), and the detection result is the voice. It is output to the recognition processing unit 21. The voice recognition processing unit 21 outputs the voice output from the voice start edge detection unit 18, the noise model corrected immediately before the utterance by the noise adaptation correction unit 19, and the acoustic model stored in the acoustic model storage unit 20 (in this case, ,
A voice recognition (S7) is performed using the noise-adapted acoustic model with low background noise, and a recognition result is obtained.

When the speech recognition processing unit 21 obtains the recognition result as described above, the data necessary for speaker adaptation (a technique for updating the acoustic model using the input user's voice), so-called teacher data. (S12) to (S1
It goes through 3). When acquiring the data necessary for speaker adaptation, a recognition score, which is represented by a score like likelihood, is used. This recognition score is given by estimating the voice uttered by the user and the content of the utterance, and if the score is higher than a predetermined threshold value, it is considered that "the recognition in the above recognition is correct" and is obtained by the recognition. This data is used as teacher data necessary for speaker adaptation.

The voice recognition processing unit 21 determines the recognition score calculated after the voice recognition (S12), and the score is higher than a predetermined threshold value (high in S12), and the request relating to the interruption of the voice recognition process by the user. Only when there is no speech feature, the recognition result after recognition and the above teacher data
It is output to the recognition result storage unit 22 and stored (S13).
This teacher data is not raw audio data, but is composed of data that is necessary for speaker adaptation, out of the features of the input voice extracted during recognition (in this example, it is called voice features). Say. The data representing the voice feature amount has a smaller data amount than the raw voice data.

On the other hand, when it is determined that the recognition score is lower than the predetermined value (low in S12), the recognition by the voice recognition is regarded as invalid as the teacher data, and the process is terminated.

Next, a processing flow in the case where the noise model B is obtained by the judgment in the noise volume judging section 13 will be described.

(Processing Flow When Adopting Noise Model B (High Noise Model)) In the noise volume determination unit 13, the noise model B
If it is determined that S. is adopted (high in S3), the noise adaptive correction unit 19 performs adaptive correction of the noise model (S9) based on the sound immediately before the user utters. At this time,
The switch in the utterance start determination switching unit 14 is the filter 1
5 and the voice storage unit 16 side. Voice input unit 1
The input voice output from 2 is separately input into two systems of a filter 15 and a voice storage unit 16, and the voice storage unit 16 stores the input voice as it is together with time information. On the other hand, the input voice input to the filter 15 is filtered by a high-pass filter and output to the pitch extraction unit 17. The pitch extraction unit 17 performs a cepstrum analysis on the input voice output from the filter 15 and extracts a portion where the pitch component starts in the input voice as a utterance start candidate (S10). In the above cepstrum analysis in this example, the point at which the correlation function of the prediction residual signal of the LPC analysis begins to exceed a certain threshold is the "point at which the sound having the pitch component begins". The pitch extraction unit 17 regards a point slightly backward from the utterance start candidate as the utterance start point, and notifies the voice storage unit 16 to output the voice from that point to the recognition processing unit. The voice recognition processing unit 21 outputs the voice output from the voice storage unit 16, the noise model corrected immediately before the utterance by the noise adaptive correction unit 19, and the acoustic model stored in the acoustic model storage unit 20 (in this case, The speech recognition (S11) is performed using the noise-adapted acoustic model with high background noise, and the recognition result is obtained.

As described above, the acoustic model stored in the acoustic model storage unit 20 is used for the speech recognition in the speech recognition processing unit 21. This acoustic model is not universal and is updated (speaker adaptation) using the input user's voice. In the voice recognition device of the present invention, the recognition result of the voice recognition processing unit 21 output from the voice feature / recognition result storage unit 22 and the voice feature amount extracted at the time of recognition (hereinafter, referred to as Speaker adaptation is performed based on the speaker adaptation voice data (abbreviated). Next, the speaker adaptation processing procedure will be described separately for pre-processing and post-processing. FIG. 3 is a flowchart for explaining preprocessing in the speaker adaptation processing, and FIG. 4 is a flowchart for explaining postprocessing in the speaker adaptation processing.

First, the preprocessing in the speaker adaptation processing will be described with reference to FIG.

(Preprocessing in Speaker Adaptation Processing) Upon receipt of the speaker adaptation voice data output from the voice feature / recognition result storage unit 22, the speaker adaptation unit 23 registers new speaker adaptation data ( Considering that the request of S21) was made,
First, it is determined whether there is a phoneme bias in the speaker adaptation voice data (S22). This determination (S22)
If it is determined that there is a bias in the number of words having the same element (S22), the word having a large bias is deleted (S23) from the point of balance with other phonemes, and the next step (S24). ). This next step (S2
In 4), it is determined whether or not the number of phonemes exceeds the threshold value (S24), and if the number of phonemes exceeds the threshold value (S2).
4)), generate a flag for speaker adaptation,
If not (not exceeded in S24), the present process is terminated. For example, assuming that the threshold is 100,
When 100 phonemes necessary for speaker adaptation have been collected, a flag for speaker adaptation is generated. This means that speaker adaptation is performed according to the number of phonemes.

On the other hand, if it is determined in the above determination (S22) that there is no bias in the number of words having the same element (No in S22), the process directly proceeds to the next step, and the number of phonemes exceeds the threshold. It is determined (S24). After this processing of (S24), processing similar to the above is performed.

Next, the post-processing in the speaker adaptation processing will be described with reference to FIG.

(Post-Processing in Speaker Adaptation Processing) The speaker adaptation unit 23 judges that the speaker adaptation start condition is satisfied (S).
In the case of 31), it is determined whether or not the speaker adaptation flag is set (S32), and if it is determined that the speaker adaptation flag is set (S32) in the determination (S32), the speaker adaptation is performed. The speaker adaptation process is executed using the voice data for use (S33), and the execution result is output to the acoustic model storage unit 20. The acoustic model storage unit 20 replaces the held speaker adaptation model with the new speaker adaptation model output from the speaker adaptation unit 23. The speaker adaptation unit 23 receives a message indicating that the speaker adaptation model has been updated, when the speaker adaptation unit 23 transmits the execution result of the speaker adaptation process to the acoustic model storage unit 20 or from the acoustic model unit 20. The stored result of the speaker adaptation process is erased (S3
4) Then, a series of speaker adaptation processing is completed. On the other hand, the above (S
32), it is determined that the speaker adaptation flag is not set (S32
If no), the speaker adaptation process is not executed and the process ends.

The speaker adaptation start condition is started at a time when it is estimated that the voice recognition device is not used for a certain time or more in one day due to the past history or the like. It is judged that the conditions are met.
In the case of a communication terminal equipped with the voice recognition device of the present invention (for example, a mobile terminal used for mobile communication), the time zone is such that the user is at home and is charging at bedtime. In such a state, the communication terminal is connected to an external power source and is being charged. That is, if a time zone in such a state is used as a speaker adaptation start trigger, the above-described speaker adaptation processing is executed while the user is sleeping and the acoustic model is automatically updated. Speech recognition using a new acoustic model will be possible without burdening the user. In addition, in the above-described state, since the communication terminal is connected to the external power supply, it is possible to always receive the power supply. Therefore, the clock speed of the control unit (CPU) in the communication terminal can be increased, and the speaker adaptation processing can be executed at high speed and completed.

As described above, according to the above-described embodiment, two noise models (noise models A and B) are prepared according to the magnitude of the background noise, and the acoustic model for each noise model is prepared. Since the voice recognition is performed by using the voice recognition, it is possible to increase the voice recognition rate in a noisy environment and save the hardware resource for the voice recognition.

In the above embodiment, it was necessary to prepare two noise models (noise models A and B) according to the magnitude of background noise and prepare acoustic models for each noise model. The embodiment of the invention is not limited to this. For example, if the original acoustic model and the noise model are held independently and can be used in combination immediately at the time of use, one acoustic model and two noise models may be held. Further, in this example, the acoustic model is generated by adapting to the magnitude of the noise volume, but the present invention is not limited to this, and may be generated by adapting to the type of noise. It does not depart from the invention.

In the above embodiment, the case where the speaker adaptation is performed using the internal expression of the corresponding voice data and the recognition result (= speaker adaptation voice data) is taken as an example. The voice data of may be stored and used as it is.

Further, in the above embodiment, the speaker adaptation is started by automatic determination (automatic determination when the voice recognition device is charging), but it may be started based on a user's instruction. For example, if it is a mobile device such as the communication terminal,
By performing a predetermined operation indicating the start of speaker adaptation, it is possible to freely perform speaker adaptation processing in a time zone desired by the user.

Further, in the above embodiment, the processing corresponding to the speaker is a local type processing mode in the voice recognition device, but the present invention is not limited to this, and the speaker adaptation is executed on the remote side. May be. In this case, the voice recognition device of the present invention is provided with a transmission unit that transmits data necessary for performing the speaker adaptation process to the outside and a reception unit that receives data after the speaker adaptation process has been executed from the outside. To be A server that executes speaker adaptation processing is provided on the remote side that is the outside. Here, a voice recognition device including the above-mentioned transmitting / receiving unit is defined as a communication terminal.

FIG. 5 is a diagram showing an example of a communication mode in the case where the speaker adaptation process is performed externally.

The case where the speaker adaptation process is performed externally will be described below with reference to FIG.

In this example, the voice feature / recognition result storage unit 2
2 includes a transmitter 5 for transmitting the stored data.
0 is connected. The voice feature / recognition result storage unit 22 sends the stored speaker adaptation voice data to the transmission unit 50 when the accumulated amount of speaker adaptation voice data necessary for speaker adaptation reaches a predetermined reference. Output. Upon receiving the speaker adaptation voice data from the voice feature / recognition result storage unit 22, the transmission unit 50 performs an encoding process on the speaker adaptation voice data, converts the speaker adaptation voice data into a wireless signal, and outputs the converted signal to a server on the remote side. (In this example, it is assumed that the notebook computer has a server function) 100. The server receives the speaker adaptation voice data transmitted from the communication terminal at the wireless transmission / reception unit 101, and executes the speaker adaptation process using the received speaker adaptation voice data. Then, after calculating a parameter to be changed (hereinafter, abbreviated as parameter change information) on the communication terminal side, the wireless transmission / reception unit 101 transmits the parameter change information to the communication terminal. Receiver 6 of communication terminal
At 0, the parameter change information transmitted from the server 100 is received and demodulated, and then the demodulated parameter change information is output to the speaker adaptation unit 23. Speaker adaptation unit 23
Then, the parameter changing process based on the parameter changing information is performed, and the update of the speaker adaptation model is completed.

As described above, according to the above-described embodiment, the speaker adaptation process is executed by the server 100 provided on the remote side, and the communication terminal side receives only the execution result from the server to obtain the speaker adaptation model. Since it is only updated, it is possible to relax computer resources and memory resources required for the communication terminal. The communication mode of the above example is the communication terminal and the server 1.
Wireless connection between 00 (eg wireless LAN, infrared communication (I
rDA), Bluetooth and other short-range wireless communication) (see FIG. 5), but a wired connection such as USB (see FIG. 5) may be used. Furthermore, a communication mode in which the server 201 on the remote side is connected to the communication network 200 as shown in FIG. 6 may be used. In this case, the communication network 200
May be a public network, a mobile network, an IP network, or the like. Also, the connection between the communication network and the communication terminal is wireless (see FIG. 6) or wired (see FIG. 6) depending on the network configuration. For example, the communication network 20
If 0 is a mobile network for mobile communication, the communication terminal wirelessly connects to a wireless access node in the mobile network, and then the server 20
Connected with 0. In the above example, the acoustic model storage unit 2
The memory function of 0 corresponds to the noise adaptive acoustic model holding unit, the voice recognition function of the voice recognition processing unit 21 corresponds to the voice recognition processing unit, and the voice detection function of the voice start edge detection unit 18 corresponds to the first voice analysis unit and the voice power. The detection means includes the filter 15, the voice storage unit 16, and the pitch extraction unit 17 for the voice detection function, the voice storage unit 16 for the voice storage function, and the pitch extraction unit 17 for utterance start. The point detection function corresponds to the pitch extraction means. Further, the switching function of the speech start determination switching unit 14 corresponds to the speech start determination switching means,
The noise correction function of the noise adaptive correction unit 19 is applied to the first noise adaptive correction unit and the second noise adaptive correction unit, and the storage function of the voice feature / recognition result storage unit 22 is applied to the voice feature recognition result storage unit, and speaker adaptation is performed. The speaker adaptation processing function of the unit 23 corresponds to the speaker adaptation means and the external speaker adaptation means. Furthermore, the transmission unit 50
Of the data transmission function of the data batch transmission means and the data sequential transmission means, the data reception function of the reception unit 60 is the data reception means, and the processing functions of the servers 100 and 201 are the data batch processing type speaker adaptation processing means and the data- The communication function of the servers 100 and 201 corresponds to the sequential storage processing type speaker adaptation processing means, and corresponds to the data batch processing type change parameter transmission means and the data reception / storage means data-sequential storage processing type change parameter transmission means.

[0061]

As described above, according to the first aspect of the present invention.
According to the present invention described in 4, the property of noise (= background noise), for example, a plurality of acoustic models corresponding to the volume of the background noise is held in advance, and is detected when the use of the voice recognition device by the user is detected. The acoustic model corresponding to the volume of the background noise to be generated is selected from the retained acoustic models. Since the voice input by the user is recognized by using the acoustic model selected as described above, even if the mobile environment is such that the noise changes greatly as the user moves, Accordingly, good recognition performance can be obtained. Further, only two acoustic models adapted to the noise volume are prepared, and each of them is switched according to the noise environment before the user uses the speech recognition device, and the speech start processing is determined by the speech analysis processing according to the noise environment. Therefore, good recognition performance can be achieved with a low processing load. Furthermore, by performing speaker adaptation processing on a server or the like provided on the remote side, the communication terminal side can reduce the burden of speaker adaptation processing, and the computer resources and memory resources required by itself can be reduced. Can be alleviated.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice recognition device in the present embodiment.

FIG. 2 is a flowchart showing a processing procedure of a voice recognition process in the voice recognition device of the present invention.

FIG. 3 is a flowchart illustrating preprocessing in speaker adaptation processing.

FIG. 4 is a flowchart illustrating post-processing in speaker adaptation processing.

FIG. 5 is a diagram (No. 1) showing an example of a communication mode when the speaker adaptation process is performed externally.

FIG. 6 is a diagram (part 2) illustrating an example of a communication mode when the speaker adaptation process is performed externally.

[Explanation of symbols]

11 Usage start / end determination unit 12 Voice input section 13 Noise volume determination unit 14 Speech start determination switching unit 15 filters 16 Voice memory 17 Pitch extractor 18 Voice Start Detection Unit 19 Noise adaptive correction unit 20 Acoustic model storage 21 Speech recognition processing unit 22 Voice feature / recognition result storage 23 Speaker adaptation unit 50 transmitter 60 Receiver 100 laptop with server function 101 wireless transceiver unit 200 communication network 201 server

─────────────────────────────────────────────────── ─── Continued Front Page (51) Int.Cl. ⁷ Identification Code FI Theme Coat (Reference) G10L 15/28 G10L 3/00 531P

Claims (14)

[Claims] 1. A voice recognition device for recognizing a user's voice input in an environment of a plurality of different noises, wherein the characteristics of the user's utterance are modeled corresponding to the respective characteristics of the plurality of different noises. Noise adaptive acoustic model holding means for holding a plurality of acoustic models described above, and when the start of use of the speech recognition device is detected, the nature of the noise is detected and an acoustic model corresponding to the detected nature of the noise is detected. A voice recognition device, comprising: a voice recognition processing unit that obtains from the noise adaptive acoustic model holding unit and collates the acoustic model with the input voice of the user to perform voice recognition. 2. The speech recognition apparatus according to claim 1, wherein the acoustic model held in the noise adaptive acoustic model holding means is generated in correspondence with the magnitude of noise volume representing the magnitude of the noise. Characteristic voice recognition device. 3. The voice recognition device according to claim 1, wherein the use start detection criterion of the voice recognition device is when the voice recognition device is gripped by a user and when the voice recognition device is operated by the user. The voice recognition device is characterized in that at least one of when the voice recognition device contacts or approaches a part of the user is detected. 4. The voice recognition apparatus according to claim 1, wherein said voice recognition means, when the noise is measured and the volume of said noise is small, the acoustic model associated with said small noise amount is said. A first noise adaptive correction unit that obtains from the noise adaptive acoustic model holding unit and corrects the background noise immediately before the utterance; and if the volume of the noise is large, the acoustic model associated with the large noise amount is used as the noise. A voice recognition device, comprising: a second noise adaptive correction unit, which is acquired from the adaptive acoustic model holding unit and is corrected by a sound immediately before utterance. 5. The voice recognition device according to claim 4, wherein the voice recognition means analyzes a voice input when the measured noise volume is low, and the measured voice recognition means. A second voice analysis unit for analyzing a voice input when the noise volume is high, and the input voice analyzed by the first voice analysis unit is corrected by the first noise adaptive correction unit. And the input voice analyzed by the second voice analysis means is recognized by using the acoustic model corrected by the second noise adaptive correction means. Speech recognizer. 6. The voice recognition device according to claim 5, wherein the voice recognition means uses the first voice analysis means and the second voice analysis means, which are used to analyze the input voice, according to the magnitude of the measured noise. A speech recognition apparatus comprising a speech start determination switching unit for switching between the speech analysis unit and the speech analysis unit. 7. The voice recognition device according to claim 5, wherein the first voice analysis unit has a voice power detection unit that determines the start of utterance using power information of the input voice, and is detected. A voice recognition device characterized in that the voice of the utterance section is outputted and recognized by the voice recognition means. 8. The voice recognition device according to claim 5, wherein the second voice analysis unit divides the input voice into two and stores one voice together with time information, and the other voice. A cepstrum analysis to obtain a starting point of the pitch component obtained by the analysis as a utterance starting point candidate, and a pitch extracting means for determining a point prior to the obtained utterance starting point candidate as an utterance starting point, and A voice recognition device characterized in that the voice recognition means outputs and recognizes the voice in the utterance section from the determined utterance start point. 9. The voice recognition device according to claim 3, wherein the voice recognition processing means calculates a recognition score indicating a recognition accuracy as a recognition result, and the calculated recognition score is predetermined. If the recognition result exceeds the threshold value, it is considered that the recognition result is highly likely to be a correct answer, and a voice feature recognition result storage unit that stores the recognition result and the voice feature amount obtained by the recognition; A voice recognition device comprising a speaker adaptation unit that learns a voice feature peculiar to a speaker by using the recognition result stored by the result storage unit and the voice feature amount obtained by the recognition. 10. The voice recognition device according to claim 9, wherein the speaker adaptation unit is stored by the voice feature recognition result storage unit when the voice recognition device is supplied by an external power supply unit. A voice recognition device, comprising: an external speaker adaptation means for performing speaker adaptation using the recognition result and the voice feature amount obtained by the recognition. 11. A voice recognition system including a communication terminal capable of wireless or wired voice communication and a server device, wherein the communication terminal is the voice recognition device according to claim 1. When the amount of data representing the voice recognition result about the input voice obtained by the recognition device and the voice feature amount is accumulated for the application of the speaker, the data is transmitted to the server device at a time. The server device includes a data receiving unit that receives the data transmitted from the communication terminal, a data batch processing type speaker adaptation processing unit that executes speaker adaptation based on the data, The data batch processing type speaker adaptation processing means calculates a parameter to be changed in the communication terminal and sends it to the communication terminal. A voice recognition system comprising a communication means. 12. The voice recognition system according to claim 11, wherein the communication terminal sequentially sends to the server device data representing a voice recognition result regarding an input voice obtained by the voice recognition device and a voice feature amount. A data receiving / storing means for receiving and storing the data transmitted from the communication terminal, and an amount of the accumulated data required for speaker adaptation. Data for performing speaker adaptation based on the data when
Sequential storage processing type speaker adaptation processing means, and data-sequential storage processing type speaker adaptation processing means calculates data to be changed in the communication terminal and transmits the data to the communication terminal-sequential storage processing type A voice recognition system comprising a change parameter transmission means. 13. The voice recognition system according to claim 11, wherein the communication terminal transmits the predetermined data by the data batch transmission means or the data sequential transmission means to a communication network connected to the server device. Voice recognition system characterized in that it is performed to the server device via 14. The voice recognition system according to claim 13, wherein the communication terminal is a mobile communication terminal.