JPH05165490A - Voice recognizing device for unspecified speaker - Google Patents

Abstract

(57) [Summary] [Purpose] To improve the recognition rate. [Constitution] The input voice is discretely divided by the vector quantizer 14.
Vector quantization code string (VQ code string),
The VQ code string is used as a hidden Markov model
Speech recognition using Mori 17 to find the probability of the phoneme
In the device, the VQ code string of the voices of many speakers
Memory of VQ code chain probability model in each phoneme section
18 and stores it in the input voice VQ code string.
Hidden Markov model and memo in memory 17 in rate calculator 15
The probability of phoneme existence is calculated using the 18 chain probability model.
Calculate That is, the current time is the j-th VQ code c_jIs output
The probability of doing o (c_j), C_iAnd c_jThe chain probability with
(C_j｜ c_i), The VQ code at the previous time is c_iPlace
If c_jProbability of occurrence of o (c_j｜ c_i) To p (c_j｜
c _i) O (c_j) / [Σp (c_m｜ c_i) O (c_m)]
Ask in. Σ is from m = 1 to N (the number of VQ codes).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for an unspecified speaker based on recognition in units of phonemes or words. A voice recognition device for an unspecified speaker, an unspecified speaker voice recognition device that does not use a speaker adaptation voice,
There is a speaker adaptation type unspecified speaker speech recognition device that uses speaker adaptation speech. An unspecified speaker voice recognition device that does not use a speaker adaptation voice is used in a system that must immediately recognize an input voice. For example, it is an automatic voice dialing system or an automatic voice ticket purchasing machine. The speaker-adaptive speaker-independent speaker recognizes some speech before the input speaker uses the system, and uses this speech to adapt the system to the input speaker. It is used, for example, in systems such as voice word processors that have limited users and allow time for the system to adapt to the speaker.

[0002]

2. Description of the Related Art In a conventional speech recognition apparatus for an unspecified speaker based on recognition on a phoneme-by-phoneme basis or a word-by-word basis, a phoneme or a word cut out from the voices of many speakers is used in advance for a short time. The acoustic parameter (for example, cepstrum or amplitude) of the voice is obtained for each, and the phoneme or word is modeled using the parameter series. Hidden Markov model (Hidden Markov model), which is one of the statistical methods, is used to model the acoustic parameter sequence of phonemes and words.
Model, hereinafter abbreviated as HMM. ; For example, there is a method using Seiichi Nakagawa). The multidimensional acoustic parameter uses a vector quantization method and a finite number (for example, 256) of discrete vector quantization codes (hereinafter referred to as VQ).
Abbreviated as code. ) Can be converted to. Therefore, the phoneme or word model is expressed using the appearance probability for each VQ code. This is called a discrete HMM. An example of the simplest one-state discrete HMM is shown in the HMM memory 7 of FIG. Here, there are four types of VQ codes (0 to 3), and phonemes (/ a /, /
The appearance probability for each VQ code is given for each a /). In practice, HMMs are prepared for all phonemes or words to be recognized. In each model, the sum of output probabilities for all VQ codes is 1.0.

FIG. 7 shows an example of the configuration of a conventional speech recognition apparatus using a discrete HMM. The voice input from the input terminal 1 is converted into a digital signal in the A / D converter 2. The acoustic parameters of this digital signal are extracted by the acoustic parameter extraction unit 3. This acoustic parameter is converted into a finite number of VQ codes in the vector quantizer 4. At this stage, the input voice is converted into the VQ code sequence. In the model probability calculation unit 5,
Each phoneme of the converted VQ code sequence or the discrete HMM of the word is read from the HMM memory 7 for the unspecified speaker, and the probability with the VQ code sequence of the input voice is calculated.

FIG. 6 shows an example of processing contents in the model probability calculating section 5. Here, it is assumed that there are four types of VQ codes (0 to 3). Given the VQ code sequence of the input voice, from time 1 to time 4 (VQ code string: 1
-1--2-0), we want to calculate the probabilities that phonemes / a / and phonemes / a / exist. The existence probability of each phoneme is
The appearance probability for the VQ code at each time is referred from the HMM of each phoneme, and is calculated by the product of them. In the example of FIG. 6, the probability for the phoneme / a / is 8.0 × 10 ^-4 ,
It is assumed that the probability for phoneme / a / is 5.0 × 10 ⁻⁵ and the probability that phoneme / a / exists is high. Such processing is performed in an arbitrary section of the VQ code string of the input voice, and the phoneme sequence that maximizes the probability is output from the recognition result output unit 6 as the recognition result.

[0005]

In the independent speaker voice recognition, a model is learned by using voice data of many speakers in order to deal with various speakers. However, as the speaker variation increases, it often happens that the distribution of acoustic parameters for one phoneme of one speaker overlaps with the distribution of acoustic parameters for different phonemes of another speaker.
For example, the phoneme / a / of speaker A is acoustically similar to the phoneme / d / of speaker B, and the type and frequency of VQ code appearing in each phoneme section are also similar. There is. This caused the recognition error.

The conventional HMM expresses only the appearance probability of the VQ code, not the temporal appearance order. FIG. 8 shows a problem that occurs because the temporal order of appearance of VQ codes is not considered. It is assumed that the VQ code sequence 1-1-0-0 is sent to the model probability calculator. In the conventional method, as shown in FIG. 8, the output probabilities for VQ code 0 and VQ code 1 are HMMs with phonemes / a /
And the phoneme / d / HMM are the same, the same appearance probability is given to the VQ code sequence. This is because the VQ code chain of 1-1 and the VQ code chain of 0-0 are very characteristic of phonemes / a /
The phoneme / d / does not have such a characteristic, and although the phoneme / a / and the phoneme / d / can be originally distinguished from each other, the information about the local VQ code chain is not stored in the phoneme HMM. Therefore, the same appearance probability is given and there is a problem that they cannot be distinguished, and the recognition performance is insufficient.

[0007]

According to the present invention, a model expressing a VQ code chain probability within a phoneme or word section is newly created and used with a conventional phoneme or word unit discrete HMM. Recognize voice by. That is, in the present invention, not only the static acoustic feature of voice but also the dynamic feature is focused. For example, speaker A
Even if the types of VQ codes appearing in different phoneme intervals of speaker B are similar, it is possible to distinguish them if the order in which they appear in time is different. The time-series pattern information is expressed by a chain probability of two or three VQ codes.

In the VQ code sequence of the input voice, the appearance probability is given from the discrete HMM to the VQ code at each time. In the present invention, the VQ code at the current time is further added.
The VQ code at the time immediately before the code, or the VQ code at the time before and immediately before the code is referred to. And
If the VQ code chain up to the current time is characteristic of the phoneme or word to be recognized, that is, if the VQ code chain probability is high, the appearance probability for the VQ code at the current time is high, and if not, it is low. To do. The VQ code chain probability is obtained in advance from the data obtained by segmenting the learning data for each phoneme. The VQ code chain probability is calculated independently for each phoneme.

If two VQ code chain probabilities are used,
The processing of the sum is expressed by an equation (1). o (c_j｜ c_i) = P (c_j｜ c_i) O (c_j) / [Σp (c_m｜ c_i) O (c_m)] (1) where o (c_j) Is the j-th VQ code c at the current time_j
Is the probability that is output, and Σ is m = 1 to N (VQ code
Up to the number of). The conventional HMM for unspecified speakers is
The recognition was performed using only the probability. p (c_j｜
c_i) Is VQ code c _iAnd VQ code c_jVQ code
It is a chain probability. This VQ code chain probability depends on the phoneme
Different. p (c_j｜ c_i) And o (c_j) Product
By, the current VQ code c_jOutput probability for
To change. The denominator of the equation (1) is all VQ at each time.
Set the sum of output probabilities for codes to 1.0
It is for. o (c_j｜ c_i) Is the VQ
The code is c_iIf it is VQ code c at the current time_jAppearance
Is the probability of doing. In this invention, the phoneme is calculated using this probability.
Recognize.

The present invention can be used in two ways depending on the data for calculating the VQ code chain probability. One is VQ
When the code chain probability is calculated from the voices of many speakers,
The other is a case where the input speaker calculates from the speaker adaptation voice uttered in advance. When the VQ code chain probability is calculated from the voices of many speakers, more general VQ code chain information of each phoneme found in many speakers is expressed in the VQ code chain probability. On the other hand, when the input speaker calculates from the speaker adaptation voice uttered in advance, the VQ code chain information peculiar to the input speaker is expressed in the VQ code chain probability. Therefore, although it is necessary for the input speaker to utter some voice before using the system, more accurate probability calculation depending on the input speaker can be performed, and the recognition performance is further improved.

[0011]

1 and 2 show the effect of utilizing the VQ code chain probability. The discrete HMM and VQ code chain probability model are created in advance from the training voice data. If the VQ code sequence 1-1-0-0 is given to the model probability calculator, the output probability for the VQ code “1” at time 2 is simply phoneme / a / and phoneme / e in the conventional discrete HMM. With reference to the occurrence probability of /, both gave a probability of 0.4. However, according to the present invention, this probability value is changed by using each VQ code chain probability of phoneme / a / and phoneme / d /. In this example, since the VQ code at time 1 immediately before time 2 is "1", the chain probability of changing from VQ code "1" to VQ code "1" is phoneme / a / and phoneme / d /. It is referred from each VQ code chain probability table, and using this, time 2
Change the probability value for the VQ code. That is, in this example, the VQ code chain of 1-1 occurs at a high probability of 0.8 in the phoneme / a / and is very characteristic. Therefore, according to the equation (1), the probability value given by the conventional discrete HMM is given. Will be appreciated more than. On the other hand, in the case of phoneme / d /, the VQ code chain of 1-1 occurs with a low probability of 0.05 and is not characteristic, so it is evaluated lower than the probability value given by the conventional discrete HMM.

As described above, by considering not only the appearance probability of VQ codes but also the chain probability of VQ codes in phonemes or words, phonemes having different appearance orders even if the appearance probabilities of VQ codes are different. Since they are recognized separately, the recognition performance is further improved.

[0013]

FIG. 2 shows an embodiment of the invention according to claim 1. The voice input from the input terminal 11 is converted into a digital signal in the A / D converter 12. An acoustic parameter (for example, cepstrum or amplitude) of this digital signal is extracted by the acoustic parameter extraction unit 13. This acoustic parameter is a finite number of Vs in the vector quantizer 14.
Converted to Q code. The number of VQ codes is, for example, 25
There are six.

After converting the voices of many speakers into the VQ code in advance, a VQ code sequence is cut out for each phoneme or word, and a phoneme or a phoneme is obtained by using an HMM represented by some number of states (for example, three states). Create a word model. These models are based on the HMM memory 17 for unspecified speakers.
Remember. Further, the chain probability of the VQ code of two or three sets is calculated from the sequence in which the voices of the above-mentioned many speakers are converted into the VQ code. This model is stored in the VQ code chain probability model memory 18.

In the model probability calculation unit 15, the HMM of a phoneme or a word is read from the HMM memory 7 for unspecified speaker, the VQ code chain probability model is read from the VQ code chain probability model memory 18, and the input voice of the input voice is read. While matching with the VQ code series, the model giving the maximum probability is examined. In this way, the recognition result output unit 16 uses the model phoneme or word that gives the maximum probability as the recognition result.
Output from.

FIG. 3 shows an embodiment of the invention of claim 2,
The parts corresponding to those in FIG. 2 are designated by the same reference numerals. As described with reference to FIG. 2, the phoneme or word HMM is created in advance using the voices of many speakers, and the HMM memory 1 for unspecified speakers is created.
Remember in 7. The present invention is applied to a speaker adaptation type apparatus, and first, the speaker adaptation mode voice recognition mode changeover switch 20 is set to the speaker adaptation side,
Ask the input speaker to speak the speaker adaptation voice. The processing up to the vector quantizer 14 is the same as in the case of FIG.
From the sequence obtained by converting this voice into a VQ code, the speaker-adapted voice VQ code chain probability calculation unit 18 calculates the chain probability of a VQ code of a set of two or three. At this time, a process for detecting a phoneme boundary is applied to the speaker-adapted voice,
The VQ code chain probability may be calculated for each phoneme and expressed by a plurality of models, or one model may be created for the entire speaker-adapted speech without detecting phoneme boundaries. The former is
It holds VQ code chain information for each phoneme of the input speaker, and the latter has VV code peculiar to the input speaker that does not consider phonemes.
Holds Q code chain information. Of course, it is better to calculate the VQ code chain probability for each phoneme because the information can be stored more precisely. These models are stored in the speaker stochastic voice VQ code chain probability model memory 19. However, when the VQ code chain probability is calculated for each phoneme, it is necessary to detect the phoneme section in the speaker-adapted speech, so that the processing amount in the speaker adaptation mode increases.

Next, the speaker adaptation mode voice recognition mode changeover switch 20 is set to the voice recognition side so that the input speaker speaks a word to be recognized. Model probability calculator 15
Then, the phoneme or word HMM is converted to the HM for unspecified speakers.
The maximum probability is read from the M memory 17, and at the same time, the VQ code chain probability model of the input speaker is read from the speaker-adapted voice VQ code chain probability model memory 19 and the maximum probability is compared with the VQ code sequence of the input voice. Examine the model you give.

In this way, the phoneme or word of the model that gives the maximum probability is output from the recognition result output unit 16 as the recognition result. In the embodiment of FIG. 3, when an input speaker is made to speak a speaker adaptation voice, an HMM for an unspecified speaker is adapted to the input speaker as in the conventional speaker adaptation voice recognition device. And the adapted HM
You may use M for the calculation of the model probability calculation part 15. In this case, the chain probability model also needs to be changed according to the change of the HMM parameters.

[0019]

As described above, in the present invention, not only the appearance probability of the VQ code in a phoneme or a word but also the chain of VQ codes is taken into consideration, so that a more precise model can be created. The recognition performance is improved. FIG. 3 shows the results of a recognition experiment of 18 consonants that have been subjected to unspecified speaker voice recognition according to the configuration shown in FIG. The voice data used is 5240 words uttered by 20 announcers (10 men and 10 women). First, 16 people (8 men
HMMs were learned for each phoneme using consonants cut out from the even-numbered words of the 5240 word set (name, 8 women). The acoustic parameters used at this time are a 16th-order cepstrum, a 16th-order delta cepstrum, and a 1st-order delta power. These are each cepstrum 25
6, Delta Cepstrum 256, Delta Power 64
It was used after being converted into VQ codes. In addition, a chain probability model of two sets of VQ codes was created for each phoneme using the same data. The time interval between two VQ codes is 8 ms
Is. The recognition experiment was conducted in 4 different people (2 men,
524 uttered by each of the two women)
It was performed using consonants cut out from the odd-numbered words of the 0-word set. FIG. 4 shows the results of an experiment for recognizing 18 consonants according to the present invention together with the conventional method. Alphabet (MT
T, MXM, FYM, FYN) are the initials of four evaluation speakers. In the conventional method, the average recognition rate is 70.
It was 8%, but was improved to 76.3% by this invention.

Next, there will be shown the results of a recognition experiment of 18 consonants which are obtained by performing speaker-adaptive unspecified speaker voice recognition according to the configuration shown in FIG. The experimental conditions are almost the same as above. However, the above 5 which were uttered by four evaluation speakers (input speakers)
216 words different from the 240 word set were used as speaker-adapted speech. From this voice, a chained probability model of two sets of VQ codes was created for each speaker. Experiments were conducted on two conditions, one for creating a VQ code chain probability model for each phoneme and one for creating one model without considering phonemes. The recognition experiment was performed using the consonants cut out from the odd-numbered words of the 5240 word set of the four evaluation speakers, similarly to the experiment using the apparatus of FIG.
FIG. 5 shows the results of the 18-consonant recognition experiment according to the present invention together with the conventional method. Alphabet (MTT, MXM, FY
M, FYN) are the initials of four evaluation speakers.
In the conventional method, the average recognition rate was 70.8%,
According to the present invention, the VQ created for each phoneme is 74.9% using the chain probability model of VQ code that does not consider phonemes.
It was improved to 78.6% using the Q-code chain probability model. When the VQ code chain probability model is created for each phoneme from the speaker-adapted speech of the input speaker, compared to the case where the VQ code chain probability model is created for each phoneme from the multi-speaker voices in FIG. Since the model is more precise for the input speaker, the performance is further improved.

[Brief description of drawings]

FIG. 1 is a diagram for explaining the principle of the present invention.

FIG. 2 is a block diagram showing an embodiment of the invention of claim 1;

FIG. 3 is a block diagram showing an embodiment of the invention of claim 2;

FIG. 4 is a diagram showing an effect of the invention of claim 1;

FIG. 5 is a diagram showing an effect of the invention of claim 2;

FIG. 6 is an HMM for a VQ code string in a conventional device
The figure which shows the example of calculation which calculates | requires the phoneme existence probability using.

FIG. 7 is a block diagram showing a conventional unspecified speaker voice recognition device.

FIG. 8 is a diagram for explaining a problem in the conventional device.

Claims (2)

[Claims] 1. An unspecified speaker is provided with a hidden Markov model memory for an unspecified speaker that stores a hidden Markov model that statistically represents a vector quantized code sequence of a phoneme or a word of speech of many speakers. An acoustic parameter vector is extracted from the input speech, the acoustic parameter vector is converted into a finite number of discrete vector quantization code sequences, and a speech recognition device for an unspecified speaker that recognizes the utterance content of the input speech, In a vector quantized code sequence of a multi-speaker, a vector quantized code chain probability model memory storing a vector quantized code chain probability model in each phoneme or each word section, and a vector quantized code sequence of the input speech On the other hand, hidden phonetic or word hidden Markov model for speaker and vector quantization code chain probability model Model probability calculation unit and, speaker independent voice recognition device, characterized in that a calculating probabilities for phonemic or word models using both. 2. An input speaker is provided with a hidden Markov model memory for an unspecified speaker that stores a hidden Markov model that statistically represents a vector quantized code sequence of a phoneme or a word of speech of many speakers. A voice recognition device for an unspecified speaker that extracts an acoustic parameter vector from speech to be recognized, converts the acoustic parameter vector into a finite number of discrete vector quantization code sequences, and recognizes the utterance content of the input speech. , A vector quantized code chain probability calculating unit for calculating a vector quantized code chain probability of a vector quantized code chain of a voice uttered by the input speaker in advance for speaker adaptation, A vector quantized code chain probability model memory storing the expressed vector quantized code chain probability model and the input speaker For the vector quantized code sequence of the voice to be recognized, a phoneme or a word using both the phoneme or word hidden Markov model for unspecified speakers and the vector quantized code chain probability model of the input speaker. A model-probability calculation unit that calculates a probability for a model, and a voice recognition device for an unspecified speaker, characterized in that.