JP2000075889A - Voice recognizing system and its method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognizing system, which is capable of following various changes in noise environment and constantly achieving a superior recognition rate, by selecting the optimal voice model in accordance with a noise environment at the time of recognition and performing a voice recognition processing by means of the voice model so selected. SOLUTION: An analytical result from a voice analyzer 11, and plural voice models 12 (voice model 1-N) are outputted to a voice recognizing part 13. The voice recognizing part 13 performs voice recognition processing independently for each voice model 1-N by referring a voice feature quantity obtained from the voice analyzer 11 to a template stored in the voice models, and outputs to a probability comparing part 14. In addition, the voice recognizing part 13 also calculates a numerical value indicating certainty other than a recognized symbol, outputting such certainty also to the probability comparing part 14. The probability comparing part 14 compares the output of the voice recognizing part 13, outputting the recognized symbol as the recognized result 15 of the system.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise-tolerant speech recognition system and a speech recognition method. For example, the present invention relates to a speech recognition system suitable for use in a car navigation system which can be operated by voice and suitable in a noise environment such as a car. And a speech recognition method.

[0002]

2. Description of the Related Art As a speech recognition technique, the classical pattern matching technique has been changed to a statistical technique in recent years, and the latter is becoming mainstream. In the latter statistical method, a Markov model having a probabilistic finite state has been proposed, and is usually referred to as an HMM (hidden Markov model: hidden Markov model). In the HMM, it is possible to increase a high recognition rate by learning a speech model using learning speech data.

FIG. 3 is a block diagram showing the configuration of a conventional continuous speech recognition system using this type of HMM.

In FIG. 3, a continuous speech recognition system comprises:
A / D conversion unit 1, LPC analysis unit 2, background noise sequential learning unit 3, voice detection unit 4, switching unit 5, Viterbi matching unit 6, HM
It comprises an M parameter estimator 7 and an HMM speech dictionary 8. The A / D converter 1 and the LPC analyzer 2
Constitutes a speech analysis block, the background noise sequential learning section 3 and the speech detection section 4 constitute a speech section detection block, and the switching section 5 and the Viterbi comparison section 6 constitute an HMM collation block.
An HMM model learning block is composed of the M parameter estimating unit 7 and the HMM speech dictionary 8.

The A / D converter 1 samples an input audio signal at a predetermined sampling frequency (for example, 8 kHz) and converts it into a digital signal.

[0006] The LPC analysis unit 2 divides a speech waveform into short sections (referred to as frames, whose length is usually 10 to 30 ms), and extracts characteristic parameters for each frame. For speech analysis, LPC (Linear Predictive Code), which is widely used as an efficient method suited to the characteristics of speech, is used.
ng: Linear Predictive Coding) analysis using LPC coefficients to LP
Calculate Cepstrum. Here, the cepstrum is obtained by performing an inverse Fourier transform of a logarithmic spectrum (Logarithm), has a property close to human auditory characteristics, and can express speech efficiently with a relatively small number of parameters.

The speech detector 4 stores the logarithmic power in the noise section and the estimated average value of the LPC cepstrum as a noise feature spectrum, finds the distance between the noise feature vector and the feature vector of the input signal, and calculates the distance from the time change. Detect voice section.

[0008] The background noise sequential learning unit 3 updates the noise feature vector in the section determined to be a noise section,
In addition to performing the sequential adaptive learning of the noise feature, the threshold is also automatically set by the adaptive learning of the distance variation.

The Viterbi collation unit 6 performs HMM collation using the Viterbi algorithm. In the HMM collation, a similarity is obtained by comparing an HMM model representing a phoneme or a word with an unknown input speech.

[0010] The HMM parameter estimating unit 7 generates an EM (Expe
HMM model learning is performed using an ctation Maximization algorithm. In the HMM model learning, the parameters of the HMM model are estimated from voice data prepared in advance.

The switching unit 5 switches the processing between the HMM collation and the HMM model learning. Also, HM
The M speech dictionary 8 stores the HM
The M model learning result is stored, and the HM
Referenced in M matching.

In general, an HMM has multiple states (eg,
And the transition between states. Furthermore, HM
M has a transition probability representing transition between states and an output probability distribution (usually using a Gaussian distribution) for outputting a feature vector accompanying the transition. FIG. 4 shows an example of word speech recognition using such an HMM.

FIG. 4 shows a word H used in the speech recognition method.
It is a state transition diagram which shows the structure of MM.

In FIG. 4, s1, s2, s3, and s4 represent states of voice features in the HMM, and a11, a12, a22, a
23, a33, a34, a44, a45 are state transition probabilities, (u1,
.sigma.1), (u2, .sigma.2), (u3, .sigma.3), and (u4, .sigma.4) represent output probability distributions.

In the HMM, the state transition probability aij (i = 1,
, 4, j = 1,..., 5)
The vector is output using the output probability distribution (uk, σk). In order to recognize an uttered word using the HMM, the HMM is first learned by using learning data prepared for each word so as to output a vector sequence of the word with the highest probability. Next, the vector sequence of the uttered unknown word is input, and the word HMM that gives the highest output probability is set as the recognition result.

In this type of speech recognition method, an uttered word itself is given an HMM and learned, and the likelihood (ie,
The recognition result is determined based on the output probability of the vector train). Although such a word HMM guarantees excellent recognition accuracy, it has disadvantages such as an enormous amount of learning data required due to an increase in the number of recognition vocabularies, and the inability to recognize speech other than the learning target words at all. is there.

In recent years, in a product such as a car navigation system, it has been attempted to perform the operation by voice from a user using voice recognition. In this case, in an environment where noise is intense, such as noise from outside the vehicle or running noise, such as inside a vehicle, it is not practical to apply speech recognition as it is without considering noise measures, as it is not practical. Therefore, various noise suppression methods have been conventionally proposed. For example,
Noise removal method called spectral subtraction, individual processing method using multiple microphones, PMC (Parallel
el Model Combination), which is an adaptation of the HMM model to noise.

At present, even with these methods, it is difficult to achieve a recognition rate that is acceptable to the user in the actual field. On the other hand, it is conceivable to perform ΗMM learning using voice data uttered in a noise environment instead of a quiet environment to create a voice model mixed with noise. This method is referred to herein as a noise learning method. By using the speech model obtained from this method, a considerable recognition rate can be achieved even in an actual field.

[0019]

However, such a conventional noise-tolerant speech recognition system has the following problems.

That is, this noise learning method has a serious restriction that the noise environment at the time of learning and the noise environment at the time of testing must be the same to achieve a good recognition rate. For example, consider an application such as voice input of a place name in car navigation. The environment surrounding the vehicle changes in various ways, such as the weather, road surface conditions, noise outside the vehicle, and traveling speed, and thus the appearance of the noise also changes.
Using a speech model trained with noise in a training speech database collected at specific weather, road surface conditions, outside noise, and traveling speed, a good recognition rate is given in the same scene as the learning conditions, but the road surface conditions If the speed changes or the running speed changes, the recognition rate will drop immediately.

As described above, it is not practical to apply the noise learning method as it is in a field where the noise environment changes variously.

The present invention provides a speech recognition system and a speech recognition method capable of always improving a good recognition rate in a field where the noise environment changes variously, even if the noise environment changes variously. The purpose is to:

[0023]

A speech recognition system according to the present invention is a speech recognition system for performing speech recognition using a speech model, comprising: a plurality of speech models prepared for different noise environments; And a voice recognition means for performing a voice recognition process using the voice model selected by the voice model selection means.

A speech recognition system according to the present invention is a speech recognition system for performing speech recognition using a speech model. In the speech recognition system, a plurality of speech models prepared for different noise environments and speech recognition using each speech model are provided. It comprises a voice recognition means for performing processing and a voice model selection means for selecting one of the recognition results recognized by the voice recognition means.

The above speech recognition processing is performed by Hidden Markov
It may be a speech recognition process using a model (HMM),
Further, the plurality of speech models may be a plurality of speech models prepared in accordance with different degrees and types of noise assumed in a field where the system is used.

Further, the plurality of speech models may be an HMM word model or an HMM phoneme model, and the plurality of speech models are noise models created by HMM learning based on speech databases uttered under different noise environments. A sound model corresponding to the environment may be used.

The voice model selecting means individually performs voice recognition processing for each voice model using the corresponding voice model, and selects a voice model given the one with the highest probability of recognition at that time. There may be.

The speech recognition system according to the present invention includes noise estimation means for estimating noise included in an input, each speech model includes a corresponding noise model, and the speech model selection means uses the noise estimation means. The estimated noise may be compared with the noise model of each speech model, and the speech model having the most similar noise model may be selected.

The matching between the estimated noise and the noise model may be performed by performing a matching process between the noise model and the estimated noise, and using the degree of matching as a measure of similarity. May be Viterbi collation.

The speech recognition method according to the present invention is a speech recognition method for performing speech recognition using a speech model.
Prepare multiple voice models for different noise environments,
Select the optimal speech model according to the noise environment at the time of recognition,
The speech recognition process is performed using the selected speech model.

A speech recognition method according to the present invention is a speech recognition method for performing speech recognition using a speech model, wherein a plurality of speech models are prepared corresponding to different noise environments, and speech recognition processing is performed using each speech model. And selecting one of the recognized recognition results.

A speech recognition method according to the present invention estimates noise included in an input, and each speech model includes a corresponding noise model. In selecting a speech model, the estimated estimated noise and each speech model are selected. And a voice model having the most similar noise model is selected.

[0033]

Embodiments of the present invention will be described below with reference to the drawings.

First Embodiment FIG. 1 is a flowchart showing the configuration and processing of a speech recognition system according to a first embodiment of the present invention.

In FIG. 1, reference numeral 10 denotes a voice input unit for converting a voice input from a microphone or the like into a digital signal and inputting it. This is a voice analysis unit including an LPC analysis unit. In the voice analysis unit 11, L which is widely used as an efficient method suitable for the characteristics of the voice is used.
The LPC cepstrum is calculated from the LPC coefficient using PC analysis.

The speech recognition system has a plurality of speech models 12 (speech models 1 to N) in advance. The speech models 12 are different speech models 1 to N corresponding to different noise environments 1 to N assumed in a field where the system is used.

The analysis result from the voice analysis unit 11 and a plurality of voice models 12 (voice models 1 to N) are output to a voice recognition unit 13 (voice recognition means).

The voice recognition unit 13 performs voice recognition processing independently for each of the voice models 1 to N by comparing the voice feature amount obtained from the voice analysis unit 11 with a template stored in the voice model, and performs a probability comparison. Output to the unit 14 (speech model selection means). Further, in the voice recognition process, the voice recognition unit 13 also calculates a numerical value indicating the probability in addition to the recognized symbol (recognized word), and also calculates the probability P-1 to PN in the probability comparison unit. 14 is output.

The probability comparing unit 14 compares the probabilities Ρ-1 to PN output from the speech recognizing unit 13, extracts the largest one Ρ-MAX, and converts the corresponding recognition symbol S-MAX into the system. Is output as the recognition result 15.

The speech recognition section 13 constitutes speech recognition means for performing speech recognition processing using each of the speech models 1 to N prepared corresponding to different noise environments, and a numerical value indicating the likelihood of the recognition result. Also, a part of the speech model selecting means is configured by calculating

The probability comparing section 14 constitutes a speech model selecting means for selecting an optimal recognition result by comparing the likelihood Ρ-1 to PN.

As described above, the speech recognition system according to the present embodiment comprises a plurality of speech models 12 (speech models 1 to N) corresponding to different degrees and types of noise, and It is characterized by having a voice recognition unit 13 that performs voice recognition processing independently and a probability comparison unit 14 that selects one of the recognition results recognized by the voice recognition unit 13.

Hereinafter, the operation of the speech recognition system configured as described above will be described.

The audio input unit 10 receives a sound from a microphone or the like and converts the signal into a digital signal by A / D conversion.

The voice analysis unit 11 performs, for example, an LPC analysis on the signal from the voice input unit 10 to extract the characteristic amount thereof, which is input to the voice recognition unit 13 at the subsequent stage.

The system has a plurality of speech models 1 in advance.
2 (voice models 1 to N). Here, the speech model may be an HMM word model or an HMM phoneme model. In order to cope with different noise environments 1 to N assumed in a field where the system is used, corresponding different speech models 1 to N are prepared.

As a method of creating these voice models,
If possible, under various noise environments in the actual field where the system is used (different SNR, type of noise, etc.)
It is desirable to create a speech database uttered in step (1) and obtain a speech model corresponding to the noise environment by HMM learning for each identical or similar noise environment. However, if it is difficult, only the noise data is obtained, and it is added to the voice database uttered in a quiet environment on a computer, and the SN ratio is changed or the type of the noise data is changed. Alternatively, a speech database may be created under various noise environments, and then a speech model may be obtained by HMM learning.

The voice recognition section 13 recognizes voice by comparing the voice feature amount obtained from the voice analysis section 11 with a template stored in the voice model. In a normal speech recognition system, one speech model is used. However, in the present embodiment, in order to cope with different noise environments, the plurality of speech models 12 (speech models 1 to N) described above are used.
Use Then, the voice recognition unit 13 performs voice recognition processing independently for each of the voice models 1 to N. In speech recognition processing based on HMM or the like, a numerical value indicating the likelihood is calculated in addition to the recognized symbol (recognized word). Therefore, the voice recognition unit 13 outputs the recognition symbols S-1 to SN when each of the voice models 1 to N is used, and also outputs the probabilities P-1 to PN at the same time.

The probability comparing section 14 compares the likelihoods Ρ-1 to PN output from the speech recognizing section 14, extracts the largest one Ρ-MAX, and extracts the recognition symbol S corresponding to the Ρ-MAX. -MAX is the system recognition result 1
Output as 5.

As described above, the speech recognition system according to the first embodiment includes different speech models 1 to N corresponding to different noise environments 1 to N assumed in a field where the system is used, and each speech model. The voice recognition process is independently performed on the models 1 to N. In the voice recognition process, a numerical value indicating the certainty is calculated in addition to the recognized symbol (recognized word).
-1 to PN are also output, and the probabilities Ρ-1 to PN which are outputs of the voice recognition unit 13 are compared with each other.
The largest Ρ-MAX is taken out, and the corresponding recognition symbol S-MAX is extracted from the system recognition result 15.
And a probability comparison unit 14 that outputs the selected speech model. The speech recognition process is performed using the selected speech model. Therefore, when speech recognition is performed, the recognition process is performed for each speech model. The recognition result of the voice model that outputs the certainty will be adopted, so that even in a situation where the noise environment changes every moment like in a car, it is possible to consistently achieve a high recognition rate following the environment. it can.

In the speech recognition system according to the first embodiment described above, a plurality of speech models 12 are prepared corresponding to different noise environments, speech recognition processing is performed using each speech model 12, and the recognized recognition result is obtained. Is selected, but instead of this, first, the voice model 12
May be selected, and speech recognition processing may be performed using the selected speech model. Hereinafter, this example will be described as a second embodiment.

Second Embodiment FIG. 2 is a flowchart showing the configuration and processing of a speech recognition system according to a second embodiment of the present invention. In the description of the speech recognition system according to the present embodiment, the same parts as those in the configuration and processing of the speech recognition system shown in FIG.

In FIG. 2, reference numeral 10 denotes a voice input unit for converting a voice input from a microphone or the like into a digital signal and inputting it. Reference numeral 11 divides a voice waveform into short sections, extracts characteristic parameters for each frame, and analyzes the voice. This is a voice analysis unit including an LPC analysis unit. In the voice analysis unit 11, the first
As in the embodiment, the LPC cepstrum is calculated from the LPC coefficient using the LPC analysis widely used as an efficient method suitable for the characteristics of the voice.

The noise estimating section 13 (noise estimating means) estimates the noise contained in the input signal and outputs an estimated noise 22.

Further, similarly to the first embodiment, a plurality of voice models 23 (voice models 1 to N) are provided in advance.
The voice models 23 are different voice models 1 to N corresponding to different noise environments 1 to N assumed in a field where the system is used, and further, each voice model 23 (voice models 1 to N). In addition, the noise model 24 of the noise environment at the time of learning corresponding to the speech models 1 to N
(Noise models 1 to N) are also prepared.

The speech model selection unit 25 (speech model selection means) compares the estimated noise 22 output from the noise estimation unit 21 with the noise model 24 stored in each speech model 23 among the speech models 1 to N. Then, the speech model having the most similar noise model is selected.

The speech recognition section 26 is composed of a speech model selection section 25
The HMM recognition processing is performed using the speech model selected by the above, and is output as a recognition result 27.

The operation of the speech recognition system configured as described above will be described below.

The voice input unit 10 receives a sound from a microphone or the like and converts the signal into a digital signal by A / D conversion. The voice analysis unit 11 performs, for example, LPC analysis on the signal from the voice input unit 10 and extracts the feature amount thereof.
Output to the noise estimation unit 21 and the speech recognition unit 26.

The noise estimator 21 estimates the noise contained in the input signal. Conventionally, various methods have been proposed as estimating methods, and these may be used. For example, a speech section in a signal may be identified, and a signal in a non-speech section may be regarded as noise.

Although the speech model 23 (speech models 1 to N) has been described in the first embodiment, here, the noise model 24 (noise model) of the noise environment at the time of learning is further provided for each speech model 1 to N. 1 to N) are also prepared separately. As a method for creating the noise model 24, when a speech database actually uttered in a noisy environment is used, a signal in a non-speech section can be regarded as noise and can be obtained by HMM learning. In the case of using the speech database created simulated above, the speech data may be obtained by HMM learning from the original noise data. Thus, the noise models 1 to N are created corresponding to the speech models 1 to N.

The voice model selecting section 25 outputs the voice model 1
Among the noise models, the estimated noise 22 output from the noise estimating unit 21 is compared with the noise model 24 stored in each audio model 23, and the audio model having the most similar noise model is selected. As the similarity measure, a matching with the estimated noise is performed by a conventional matching method (Viterbi matching or the like) using a noise model, and a matching degree (probability in HMM) output as a result is used. When a one-state, one-mixture HMM model is selected as the noise model, the noise model is expressed only by the mean and the variance. Therefore, the mean and the variance are calculated in advance from the estimated noise, and the difference between the estimated noise and the noise model is calculated. Gaussian distance may be used as a similarity measure.

In the speech recognition section 26, the speech model selection section 2
The HMM recognition process is performed using the speech model selected in Step 5, and a recognition result 27 is output, and the process ends. In this case, there is no need to perform recognition processing on a plurality of speech models as in the first embodiment.

As described above, the speech recognition system according to the second embodiment further includes, for each of the speech models 23 (speech models 1 to N), the noise at the time of learning corresponding to the speech models 1 to N. Environmental noise model 24 (noise models 1 to
N) is also prepared, and among the speech models 1 to N, the estimated noise 22 output from the noise estimating unit 21 is compared with the noise model 24 stored in each speech model, and the most similar noise model is obtained. Voice model selection unit 25 for selecting a voice model
And a speech recognition unit 26 that performs HMM recognition processing using the speech model selected by the speech model selection unit 25, estimates a noise component from an input, and is most similar to the estimated noise among a plurality of speech models. Since a speech model having the noise model obtained in advance is selected and recognition processing is performed only on the speech model, the processing load of the speech recognition unit 26 is significantly reduced as compared with the first embodiment. Real-time speech recognition becomes possible.

Therefore, if the speech recognition system having such excellent features is applied to, for example, a noise-tolerant speech recognition system used for car navigation which can be operated by voice, the recognition rate can be greatly improved in this device. Can be planned.

The speech recognition system and the speech recognition method according to each of the above embodiments can be applied to any speech recognition system that performs speech recognition using a speech model.
The present invention can be implemented as a circuit incorporated in various terminals.

Further, in each of the above embodiments, the plurality of speech models correspond to speech models created by HMM learning, in particular, noise environments created by HMM learning based on speech databases uttered under different noise environments. It is desirable to use a voice model, but any method may be used as long as a plurality of voice models are used according to the degree and type of noise.

Further, the number of processing units and various processes constituting the speech recognition system according to each of the above embodiments, the type of connection, and the like are not limited to the above embodiments.

[0069]

According to the speech recognition system and the speech recognition method of the present invention, a plurality of speech models prepared corresponding to different noise environments, speech recognition means for performing speech recognition processing using each speech model, Since it has a voice model selecting means for selecting one of the recognition results recognized by the voice recognizing means, in a field where the noise environment changes variously, even if the noise environment changes variously, it follows the noise environment. A good recognition rate can always be raised.

In the speech recognition system and the speech recognition method according to the present invention, a plurality of speech models prepared corresponding to different noise environments and a speech model selection for selecting an optimal speech model according to the noise environment at the time of recognition. Means and voice recognition means for performing voice recognition processing using the voice model selected by the voice model selection means, it is possible to increase the recognition rate while significantly reducing the load of voice recognition processing, Real-time speech recognition becomes possible.

[Brief description of the drawings]

FIG. 1 is a flowchart showing the configuration and processing of a speech recognition system according to a first embodiment to which the present invention has been applied.

FIG. 2 is a flowchart showing the configuration and processing of a speech recognition system according to a second embodiment to which the present invention has been applied.

FIG. 3 is a block diagram showing a configuration of a conventional continuous speech recognition system using an HMM.

FIG. 4 is a diagram showing a structure of a word Hidden Markov Model used in a speech recognition method.

[Explanation of symbols]

Reference Signs List 10 voice input unit, 11 voice analysis unit, 12, 23 multiple voice models (voice models 1 to N), 13 voice recognition unit (voice recognition unit), 14 probability comparison unit (voice model selection unit), 15 recognition result, 21 noise estimation unit (noise estimation means), 22 estimated noise, 24 noise model, 25 speech model selection unit (speech model selection means), 26 speech recognition unit, 27 recognition result

Claims (13)

[Claims] 1. A speech recognition system for performing speech recognition using a speech model, wherein a plurality of speech models prepared for different noise environments and an optimal speech model are selected according to the noise environment at the time of recognition. A voice model selecting means for performing a voice recognition process using the voice model selected by the voice model selecting means. 2. A speech recognition system for performing speech recognition using a speech model, comprising: a plurality of speech models prepared corresponding to different noise environments; and speech recognition means for performing speech recognition processing using each speech model. And a voice model selecting means for selecting one of the recognition results recognized by the voice recognizing means. 3. The speech recognition system according to claim 1, wherein the speech recognition process is a speech recognition process based on a Hidden Markov Model (HMM). 4. The voice model according to claim 1, wherein the voice models are voice models prepared according to different noise levels and types assumed in a field where the system is used. 3. The speech recognition system according to any one of 2. 5. The speech recognition system according to claim 1, wherein the plurality of speech models are HMM word models or HMM phoneme models. 6. The plurality of speech models are based on speech databases uttered under different noise environments.
The speech recognition system according to claim 1, wherein the speech model is a speech model corresponding to a noise environment created by MM learning. 7. The voice recognition means performs voice recognition processing individually for each voice model, and the voice model selection means selects a voice model given the highest probability of recognition in the voice recognition processing. 3. The speech recognition system according to claim 2, wherein said speech recognition system is selected. 8. A noise estimation unit for estimating noise included in the input, each speech model includes a corresponding noise model, and the speech model selection unit includes an estimated noise estimated by the noise estimation unit. 2. The speech recognition system according to claim 1, wherein the speech model is compared with a noise model of each speech model, and a speech model having the most similar noise model is selected. 9. The collation between the estimated noise and the noise model is as follows:
Perform a matching process between the noise model and the estimated noise,
9. The speech recognition system according to claim 8, wherein the degree of matching is used as a measure of similarity. 10. The speech recognition system according to claim 9, wherein said matching processing is Viterbi matching. 11. A speech recognition method for performing speech recognition using a speech model, comprising preparing a plurality of speech models corresponding to different noise environments, and selecting an optimal speech model according to the noise environment at the time of recognition. And performing a voice recognition process using the selected voice model. 12. A speech recognition method for performing speech recognition using a speech model, comprising: preparing a plurality of speech models corresponding to different noise environments; performing speech recognition processing using each speech model; A voice recognition method comprising selecting one of the following. 13. Estimating noise included in an input, wherein each speech model includes a corresponding noise model. In selecting the speech model, the estimated noise and the noise model of each speech model are combined. The speech recognition method according to claim 11, wherein the speech model having the most similar noise model is selected by collation.