JP2980382B2 - Speaker adaptive speech recognition method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speaker adaptive speech recognition method and apparatus for generating and using a conversion function for converting a feature amount of a new speaker's speech into a feature amount of a standard speaker's speech. The voice recognition device recognizes a voice that has been uttered by comparing a voice feature amount registered in advance with a voice feature amount extracted from an input voice. If the pre-registered speech features are created by the same person as the input speech to be recognized, the recognition accuracy will be high,
If it is made standard for an unspecified speaker, the recognition accuracy will be reduced especially for highly individualized utterances.

However, creating a dictionary of speech features for a specific (new) speaker requires a great deal of labor. Therefore, a dictionary indicating the standard speech features is created in advance, and a conversion function that converts the speech features into the standard speaker speech features for new speaker speech is generated by learning. Then, using the conversion function, the feature of the input speech is converted into a form close to the feature of the standard speaker's speech, and then the matching is performed. There is a need for a technique for easily and quickly generating a conversion function used at this time.

[0003]

2. Description of the Related Art FIG. 7 is an explanatory diagram of the prior art. When generating the conversion function 60, the feature time series used for speech recognition is extracted from the speech of the new speaker by the new speaker speech feature extraction means 1, and the standard speech corresponding to the speech of the new speaker is extracted. The feature time series of the speaker's voice is stored in a standard speaker voice feature storage unit 4.
, And a conversion function generating unit 30 generates a single conversion function 60 for converting the feature time series of the new speaker's voice into the feature time series of the standard speaker's voice.

[0004] When performing speech recognition with the voice of a new speaker,
The feature amount time series used for speech recognition is extracted from the new speaker's speech by the new speaker speech feature amount extraction means 1, and the feature amount is converted into the feature amount time series of the standard speaker voice by the conversion function 60. The feature time series and the feature time series stored in the standard speaker voice feature storage unit 4 are collated by the standard speaker voice recognition unit 7 to obtain a recognition result.

[0005]

According to the prior art described above, all new speaker voice features are converted to standard speaker voice features by a single conversion function. Since the conversion function becomes extremely complicated and it takes a long time to generate the conversion function, there is a problem that the load on a new speaker is large. Also, there is a problem that it is difficult to obtain a conversion function with good conversion accuracy.

SUMMARY OF THE INVENTION It is an object of the present invention to solve the above problems, to provide a means for generating an accurate conversion function in a short time, and to reduce the burden on a new speaker.

[0007]

FIG. 1 is a block diagram showing the principle of the present invention. According to the present invention, when the feature amount of the new speaker's voice is converted into the feature amount of the standard speaker's voice by the conversion function, the conversion function corresponding to the acoustic characteristic of the new speaker's voice at each time from among the plurality of conversion functions. Is selected, and the feature amount of the new speaker's voice is converted into the feature amount of the standard speaker's voice while switching the conversion function used for the conversion.

[0008] N conversion functions corresponding to N kinds of acoustic characteristics
When generating a transformation function group consisting of (1), (2), ... (N),
The following processing is performed. From the input new speaker's voice, a new speaker's voice feature amount extracting unit 1 extracts a feature time series of the new speaker's voice. The acoustic characteristic extraction unit 2 extracts the acoustic characteristic time series corresponding to the N conversion functions from the feature amount time series.

The conversion function generating means 3 reads out the characteristic time series of the standard speaker voice corresponding to the input new speaker voice from the standard speaker voice characteristic quantity storing means 4 and corresponds to the acoustic characteristic at each time. Then, a conversion function group including N conversion functions for converting the feature amount of the new speaker's voice into the feature amount of the standard speaker's voice is generated.

[0010] When performing speech recognition with the voice of a new speaker,
The following processing is performed. The feature time series of the new speaker's voice is extracted by the new speaker's voice feature extracting means 1 from the voice of the new speaker. The acoustic characteristic extraction unit 2 extracts the acoustic characteristic time series corresponding to the N conversion functions from the feature amount time series.

The conversion function switching means 5 selects a conversion function corresponding to the acoustic characteristic at each time from a group of N conversion functions. The speech feature amount conversion means 6 converts the input feature amount of the new speaker's speech into the feature amount of the standard speaker's speech using the selected conversion function. That is, the feature time series of the new speaker's voice is converted into the feature time series of the standard speaker's voice by the conversion function group.

[0012] The standard speaker voice recognition means 7 includes a feature time series of the standard speaker voice read out from the standard speaker voice feature storage means 4, and a new speaker voice feature time sequence obtained by the voice feature conversion means 6. A feature amount time series obtained by converting the series is collated and a recognition result is output.

As the acoustic characteristic extracted by the acoustic characteristic extracting means 2, a distinctive feature can be used. The discrimination feature indicates an acoustic property such as vowel, consonant, monoformant, nasal, energetic, and so on.

Further, specifically, properties such as voiced sound, unvoiced sound, and silence may be selected as acoustic characteristics, and a conversion function may be prepared for each of these properties.

The conversion function generating means 3 can obtain a conversion function for each acoustic characteristic by regression analysis. Also,
The conversion function generation means 3 obtains the conversion function by a neural network, and the speech feature quantity conversion means 6 can also convert the feature quantity of the new speaker's voice by the conversion function realized by the neural network corresponding to the acoustic characteristics. it can.

[0016]

According to the present invention, a plurality of conversion functions corresponding to the acoustic characteristics of the input speech are generated, and these functions are switched and used at the time of recognition according to the acoustic characteristics of the input speech at each time. Since the features of the speaker's speech are converted to the features of the standard speaker's speech, each conversion function becomes simple. For example, learning with 100 words of speech input was required to generate the conversion function. On the other hand, even if learning is performed by inputting a very small number of words, an accurate conversion function can be realized.

[0017]

FIG. 2 shows an embodiment of the present invention. In FIG. 2, the band spectrum time series calculation unit 11 corresponds to the new speaker voice feature amount extraction unit 1 shown in FIG. The voiced / unvoiced / silent determining unit 21 is a sound characteristic extracting unit 2 shown in FIG.
Corresponding to The DP matching unit 31, the band spectrum pair storage selection unit 32, the band spectrum pair storage unit group 33 for each acoustic characteristic, and the linear regression analysis unit 34 correspond to the conversion function generation unit 3 shown in FIG. The standard speaker word voice band spectrum time-series template storage unit 41 corresponds to the standard speaker voice feature storage unit 4 shown in FIG. The conversion function selector 51 corresponds to the conversion function switching means 5 shown in FIG. The voice feature value conversion unit 61 corresponds to the voice feature value conversion means 6 shown in FIG. DP matching voice recognition unit 71
Corresponds to the standard speaker voice recognition means 7 shown in FIG.

The speech feature quantity conversion unit 61 has conversion functions 61a, 61b, and 61c for voiced, unvoiced, and silent.
When generating these conversion functions, the following processing is performed.

The band spectrum time series calculation unit 11 converts a new speaker's voice into a band spectrum time series. Voiced sound
The unvoiced sound / silence determination unit 21 determines whether the acoustic characteristic at each time is voiced, unvoiced, or silent based on the band spectrum time series. Whether or not there is no sound is determined based on the magnitude of the total power of the band spectrum. Whether the sound is a voiced sound or an unvoiced sound is determined by the relative magnitude of the low band power and the high band power of the band spectrum.

The DP matching section 31 performs time matching between the input band spectrum time series of the new speaker and the band spectrum time series corresponding to the input speech in the standard speaker word speech band spectrum time series template storage section 41. Take
Generate a band spectrum pair. The band spectrum pair is a pair of the band spectrum of the new speaker and the band spectrum of the standard speaker associated by time matching.

The band spectrum pair storage selection unit 32 stores the band spectrum pair in the band spectrum pair storage unit group 33 according to the judgment of the voiced sound / unvoiced sound / silence judgment unit 21. The data is stored and stored in one of the storage unit and the silent band spectrum pair storage unit. For example, if the determination by the voiced / unvoiced / silence determining unit 21 is a voiced sound, the band spectrum pair is stored and stored in the voiced band spectrum pair storage unit.

The linear regression analysis unit 34 performs a linear regression analysis on the band spectrum pairs stored in the respective storage units in the band spectrum pair storage unit group 33 to obtain each conversion function, and obtains a speech feature amount conversion unit 61. Notify and store it. For example, by performing a linear regression analysis on the band spectrum pair stored in the voiced sound band spectrum pair storage unit, a conversion function when the input voice is a voiced sound is generated, and the voice feature amount conversion unit 61 manages the conversion function. Conversion function 6 for voiced in the conversion function group
1a.

When performing speech recognition with the voice of a new speaker,
The following processing is performed. Band spectrum time series calculator 11
Converts the new speaker's speech into a band spectrum time series.
The voiced / unvoiced / silence determining unit 21 determines whether the acoustic characteristic at each time is voiced, unvoiced, or silent based on the band spectrum time series. According to this determination, the conversion function selection unit 51 selects one of the conversion functions from the conversion function group used by the audio feature value conversion unit 61.

At a certain time, for example, if the voiced / unvoiced / silent determining unit 21 determines that the voice is a voiced sound, the conversion function selecting unit 51 converts the band spectrum of the new speaker's voice at that time into the standard speaker's voice. The voiced conversion function 61a is selected as a conversion function for converting to a band spectrum.

The speech feature quantity converter 61 converts the band spectrum time series of the new speaker's voice into the band spectrum time series of the standard speaker's voice according to the selected conversion function. The DP matching speech recognition unit 71 collates the converted band spectrum time series with the band spectrum time series for each word in the standard speaker word speech band spectrum time series template storage unit 41, and outputs a recognition result.

FIG. 3 shows a processing flow at the time of generating a conversion function, that is, a processing flow at the time of learning according to the embodiment shown in FIG. Hereinafter, description will be made in accordance with the processes (a) to (l) shown in FIG.

(A) The words spoken by a new speaker for learning are predetermined. When the new speaker utters, the band spectrum time series calculation unit 11 converts the new speaker waveform of the voice into a band spectrum time series. (b) The DP matching unit 31 performs DP matching between the new speaker band spectrum time series and the standard speaker band spectrum time series of the same word to generate a band spectrum pair.

(C) Processes (d) to (i) are repeated until the time series ends. When the processing is completed, the processing moves to the processing (j). (d) The voiced / unvoiced / silence determining unit 21 determines whether the band spectrum is voiced / unvoiced / silent. (e) to (f) If it is a voiced sound, the band spectrum pair is stored in the voiced sound band spectrum pair storage unit in the band spectrum pair storage unit group 33. (g) to (h) If the sound is unvoiced, the band spectrum pair is stored in the unvoiced band spectrum pair storage unit in the band spectrum pair storage unit group 33. (i) If there is no sound, the band spectrum pair is stored in the silent band spectrum pair storage unit in the band spectrum pair storage unit group 33. Thereafter, the process returns to the process (c), and the process is repeated in the same manner.

(J) When the time series is completed, a silence conversion function 61c is generated by linear regression analysis and stored. In addition,
Since the method of linear regression analysis is well known, a detailed description thereof will be omitted. (k) Similarly, a conversion function 61b for unvoiced sound is generated by linear regression analysis and stored. (l) Similarly, a voiced sound conversion function 61a is generated by linear regression analysis and stored. With the above processing, the conversion function generation processing ends.

FIG. 4 shows a processing flow at the time of recognition according to the embodiment shown in FIG. Hereinafter, the processing (a) to FIG.
Explanation will be given according to (k).

(A) The speech uttered by the new speaker is converted into a band spectrum time series by the band spectrum time series calculation unit 11.

(B) Processes (c) to (i) are repeated until the time series ends. When the processing is completed, the processing moves to the processing (j). (c) The voiced / unvoiced / silent determining unit 21 determines whether the band spectrum is voiced / unvoiced / silent. (d) to (e) If it is a voiced sound, select a voiced conversion function as the conversion function. (f) to (g) If it is an unvoiced sound, select a conversion function for unvoiced as a conversion function. (h) If there is no sound, select the silence conversion function as the conversion function. (i) The speech feature conversion unit 61 converts the band spectrum into that of the standard speaker by using the selected conversion function. Thereafter, the process returns to the process (b), and the process is repeated in the same manner.

(J) When the time series of the recognition target is completed,
The DP matching voice recognition unit 71 performs DP (dynamic programming) matching on the converted band spectrum time series and the band spectrum time series template of the standard speaker. (k) As a result of the DP matching, the phrase having the best score is regarded as a recognition result, and the processing is terminated.

The conversion function group is composed of a plurality of neural networks to form a neural network group.
Can be realized as a neural network learning unit by back propagation.
FIG. 5 shows an embodiment of the present invention using the neural network.

In FIG. 5, the same reference numerals as those in FIG.
Correspond to those shown in FIG. 35 is a back propagation unit for learning the neural network, 52 is a neural network selecting unit for selecting a neural network to be used for conversion by judging voiced sound, unvoiced sound, or no sound, and 62 is a conversion for converting a band spectrum by the neural network. , 62a represents a voiced neural network, 62b represents a voiceless neural network, and 62c represents a silent neural network.

In the embodiment shown in FIG. 5, when generating a conversion function group, that is, a voiced neural network 62a, a voiceless neural network 62b, and a silent neural network 62c, the voiced sound The processing until the band spectrum pairs are separately stored in the storage unit provided for each unvoiced sound / silence is the same as in the embodiment of FIG.

In this embodiment, learning is performed by the back propagation unit 35 in order to realize the conversion function by a neural network. Here, a band spectrum pair is read from the band spectrum pair storage unit group 33 for each storage type, and for example, a band spectrum of a new speaker's voice is used as an input signal to be supplied to a neural network, and a band spectrum of a standard speaker's voice is used as a teacher signal. In this way, learning to determine the internal state of each neuron is performed.

In general, a neural network is often composed of neurons arranged in an input layer, an intermediate layer, and an output layer. In the case of the present embodiment, however, since the transformation function is simplified, it is substantially realized. Even in linear conversion, considerable conversion accuracy can be maintained. Therefore, as shown in FIG. 6, the intermediate layer may be omitted, and a neural network including only the input layer and the output layer may be used. The silence neural network 62c may transmit the input signal of the input layer to the output layer as it is.

Each neural net 62a,
When 62b and 62c are created, speech recognition using them is performed as follows.

Referring to FIG. 5, a voiced / unvoiced / silent determining unit 21 determines a time series of a band spectrum of a new speaker's voice.
The acoustic characteristics at each time are voiced, unvoiced,
The processing up to the determination of silence is the same as in the embodiment shown in FIG.

The neural network selecting section 52 outputs a voiced sound
Based on the judgment result of the unvoiced sound / silence judgment unit 21, the conversion unit 62
Select the neural net to use in. That is, as shown in FIG. 6, the voiced neural network 62a, the unvoiced neural network 62b,
One of the three types of neural networks 2c and 2c is selected, and a new speaker band spectrum time series is used as an input signal to the selected neural network. With this input, the signal output from the output layer corresponds to the time series of the band spectrum of the standard speaker voice.

The band spectrum time series converted by the neural network is passed to the DP matching speech recognition unit 71. Subsequent speech recognition processing is the same as in the embodiment shown in FIG.

In the above embodiments, examples of voiced sound, unvoiced sound, and silence have been taken as the acoustic characteristics. However, the present invention is not limited to this, and can be similarly implemented using various discrimination features.

[0044]

As described above, according to the present invention,
Since a plurality of conversion functions corresponding to the acoustic characteristics are used, each conversion function is simplified, and the conversion function can be generated in a short time. Therefore, the burden on new speakers is reduced. In addition, the conversion accuracy is improved, and a good recognition result can be obtained.

[Brief description of the drawings]

FIG. 1 is a principle block diagram of the present invention.

FIG. 2 is an explanatory view of an embodiment of the present invention.

FIG. 3 is a diagram showing a processing flow at the time of generating a conversion function according to the embodiment of the present invention.

FIG. 4 is a diagram showing a processing flow at the time of recognition according to the embodiment of the present invention.

FIG. 5 is an explanatory view of an embodiment using the neural network of the present invention.

FIG. 6 is a diagram illustrating an example of a neural network according to an embodiment of the present invention.

FIG. 7 is an explanatory diagram of a conventional technique.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 New speaker voice feature extraction means 2 Acoustic characteristic extraction means 3 Conversion function generation means 4 Standard speaker voice feature storage means 5 Conversion function switching means 6 Voice feature conversion means 7 Standard speaker speech recognition means

──────────────────────────────────────────────────続 き Continued on the front page (58) Fields surveyed (Int. Cl. ⁶ , DB name) G10L 3/02 301 G10L 3/00 521 G10L 3/00 531 G10L 3/00 539 G10L 9/10 301

Claims (6)

(57) [Claims] 1. A speaker adaptive speech recognition method for generating and using a conversion function for converting a feature amount of a new speaker's speech into a feature amount of a standard speaker's speech. The conversion function used in a plurality of conversion functions is switched based on the process of generating the conversion function by learning and the acoustic characteristics at each time of the input speech during recognition, and the feature amount of the new speaker's voice is changed to the standard speaker. Converting to a speech feature quantity. 2. The speaker adaptive speech recognition method according to claim 1, wherein a discrimination feature is used as an acoustic characteristic corresponding to each of the plurality of conversion functions. 3. The speaker adaptive speech recognition method according to claim 1, wherein a voiced sound, an unvoiced sound, and a silent sound are used as acoustic characteristics corresponding to each of the plurality of conversion functions. 4. The speaker adaptive speech recognition method according to claim 1, wherein the conversion function is obtained by regression analysis. 5. The speaker adaptive speech recognition method according to claim 1, wherein the conversion function is obtained by a neural network, and the plurality of conversion functions realized by the neural network corresponding to the acoustic characteristics are obtained. A speaker-adaptive speech recognition method characterized by converting a feature amount of an input speech into a feature amount of a standard speaker speech. 6. A speaker adaptive speech recognition apparatus for performing speech recognition by generating and using a conversion function for converting a feature amount of a new speaker's speech into a feature amount of a standard speaker's speech. New speaker's voice feature extraction means (1) for extracting the feature amount from the speech of the user, and acoustic characteristic extraction means (2) for extracting a sound characteristic time series related to a plurality of predetermined acoustic characteristics from the extracted feature amount. ), Standard speaker voice feature storage means (4) that stores the feature time series of the standard speaker voice to be referenced during recognition, and the feature volume of the standard speaker voice corresponding to the input new speaker voice A plurality of conversion functions for reading the time series from the standard speaker voice feature storage means (4) and converting the feature of the new speaker voice into the feature of the standard speaker voice corresponding to the acoustic characteristics at each time. The acoustic characteristic extracting means
A conversion function generating means (3) for generating according to the acoustic characteristics extracted by (2);
The conversion function switching means (5) for switching the conversion function to be used in accordance with the acoustic characteristics at each time of the input voice extracted in (2), and the conversion function selected by the conversion function switching means (5), the new speaker Speech feature conversion means (6) for converting speech features into standard speaker speech features
And a standard speaker speech recognition unit (7) for performing speech recognition by comparing the feature amount of the converted speech with the feature amount of the standard speech read from the standard speaker speech feature amount storage unit (4). A speaker adaptive speech recognition apparatus, comprising: