JPH1195789A - Voice recognition system and speaker adaptive method in the same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable anyone to use a voice recognition system effectively by performing a voice recognition while using the acoustic model of the voice recognition system suited for the feature of a speaker estimated in accordance with a detected physical feature information of the speaker to compensating the degradation of a recognition rate to be caused by depending on feature of the voice of a user. SOLUTION: The feature of the user is extracted from a voice 100 in the personal information extracting part 310 of the voice of the user by using the voice 100 or measured results of a mat type scale 160 and/or a sheet position detecting meter 170 as input information. Moreover, when the measured information of the mat type scale 160 and/or the sheet potion detecting meter are used as the input information, the information are converted into information of a weight or a stature in a physical feature converting part 320 and these are converted into the feature of the voice of the user in a voice feature conversion part 330. Then, a speaker adaptation is executed in a speaker adaptation part 350 for acoustic model so that an acoustic model to be used in the voice recognition becomes the acoustic model suited for the user based on the personal information detected in this manner.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a speech recognition system, and more particularly to a speaker adaptation method and apparatus capable of improving a recognition rate in conformity with a speaker.

[0002]

2. Description of the Related Art FIG. 1 is a diagram showing a conventional concept of speaker adaptation which led to the invention. In the conventional speaker adaptation, using the input speech 100, a process 300 for adapting an acoustic model suitable for the speaker is performed, a conventional speech recognition process 500 is performed, and a recognition result 999 is output. The acoustic model adaptation process 300 is a process 310 for extracting personality information.
And a process 350 for adapting an existing acoustic model from the results
It is composed of

[0003]

In the conventional speech recognition system, the recognition rate is still highly dependent on the speaker, and it is not clear what features the speaker adaptation is to be performed in when the speaker adaptation is performed. There was a problem.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a speaker adaptation method capable of effectively using a speech recognition system by compensating for a deterioration of a recognition rate depending on characteristics of a user's speech by more reliable speaker adaptation. It is to provide a method and an apparatus.

[0005]

In order to achieve the above object, in a speaker adaptation method in a speech recognition system according to the present invention, physical characteristic information of a speaker is detected, and the characteristic is detected in accordance with the physical characteristic information. It is characterized by estimating the characteristics of the speaker's voice and performing voice recognition using an acoustic model of a voice recognition system suitable for the estimated characteristics of the voice.

The physical characteristic information is, for example, at least one of the speaker's weight information and height information.

[0007] Such physical characteristics are basically proportional to the scale of the size and length of the voice vocal organ, and it is possible to infer the characteristics of the voice vocal organ from such information. .

According to the present invention, it is possible to perform more appropriate speaker adaptation, which cannot be conventionally performed by speaker adaptation based on a speaker's speech, and as a result, the speech of an arbitrary speaker can be improved. Can be improved.

A speech recognition system equipped with a speech recognition function according to the present invention as a system for realizing such a method includes means for acquiring physical characteristic information of a speaker, physical characteristic information of the speaker, and speech. A correspondence table that associates the features of the above, means for extracting the features of the speaker's voice by illuminating the acquired physical feature information with the correspondence table, and a plurality of acoustic models for speech recognition, Acoustic model adaptation means for selecting or converting an acoustic model suitable for the extracted speech feature from the plurality of acoustic models, and speaker's speech recognition using the acoustic model obtained by the acoustic model adaptation means Voice recognition processing means for performing processing.

When the voice recognition system is mounted on a car navigation system, the means for acquiring the weight information includes a mat-type weighing scale arranged on a seat of a speaker, and the means for acquiring the height information includes a speech. It has a seat position detector for detecting the position of the person's seat.

When the seat position detector cannot be used for obtaining height information as in car navigation, the means for obtaining the height information includes a video camera for obtaining an image of a speaker and processing of the image to obtain the height information. Can be constituted by processing means for obtaining

[0012] In addition to the physical features, the speech characteristics of the speaker are identified from the utterance content of the adaptation word specified in advance by the speaker, and based on the speech features and the physical feature information of the speaker. You may use it combining the characteristic of the obtained audio | voice. In this case, the acoustic model adaptation means selects or converts the acoustic model based on the characteristics of the combined speech.

At least two types of acoustic models are provided for men and women. Further, a separate acoustic model may be provided for each physical feature in advance.

If the user is provided with a selection means for selecting whether or not to apply the speaker adaptation of the acoustic model prior to the speaker's speech recognition, the speaker adaptation can be performed only when necessary. .

[0015]

Embodiments of the present invention will be described below in detail.

FIG. 2 is a conceptual diagram of a process showing a speaker adaptation method according to the present invention. The speaker adaptation process 300 includes, in addition to the conventional speaker adaptation process based on the features of speech described in FIG.
In this method, personal characteristic information of a voice is extracted from physical characteristic information such as the height and weight of a speaker characterizing the voice, and an existing acoustic model is adapted.

More specifically, the voice 100, the mat-type weight scale 160, and / or the seat position detector 170
The voice personality information extraction unit 310 extracts characteristics of the user's voice from the voice 100 using the measurement result of the above as input information. When the mat type weight scale 160 and / or the seat position detector 170 are input information,
First, in the physical feature conversion unit 320, the weight or height is converted into information, and in the voice feature conversion unit 330,
It is converted into features of the user's voice. Based on the personality information of the voice extracted by the above means, the speaker adaptation unit 350 of the acoustic model uses the speaker so that the acoustic model used for recognition becomes an acoustic model suitable for the user. Adaptation is performed. Thereafter, normal speech recognition 500 is performed, and a recognition result 999 is output.

FIG. 3 is a diagram showing an example of means for detecting a physical characteristic of a speaker, which is the main point of the present invention. In the present embodiment, an application of voice recognition to a car navigation device is described.

First, in a vehicle 1000 equipped with a car navigation device, as shown in FIG.
Below 0, a mat-type weight scale 160 is provided, and the weight of the seat occupant, usually the driver, is automatically detected. When the driver positions the seat, the approximate height of the driver is detected from the seat position detector 170 provided. These weight and height scales are basically proportional to the size and length scales of the voice vocal organs, and it is possible to infer the characteristics of the voice vocal organs from the weight and height.

FIG. 4 is a block diagram showing an example of the configuration of the present invention in detail. With the audio signal 100 as an input, the audio signal is sampled by an LPF (Low Pass Filter) 2010 and an A / D converter 2020 of the audio input unit 200, and analog audio information is converted into digital audio waveform information. You. Then, in the voice analysis unit 210,
The voice feature parameters are extracted. The speech parameters include, for example, an LPC cepstrum determined by a linear prediction analysis. For details, see, for example, the document “Basics of Speech Information Processing” (co-authored by Shozo Saito and Kazuo Nakata, Ohmsha). Then, in the personality information extraction unit 310,
Voice personal information (personality information) is extracted. In the present embodiment, the personality information is fundamental frequency (pitch) information that defines information on the pitch of the voice. The personality information obtained based on the voice signal 100 is combined with personality information 34 obtained from physical characteristics described later.
0 is combined. In this merging process 340,
For example, a method of averaging personality information obtained from voice and personality information obtained from physical characteristics with a certain weight is adopted. Alternatively, two pieces of information are merged using a certain conversion table. Specifically, when viewed from the pitch information, the value is the value of the pitch information resulting from the merging.

In the present invention, weight and height are taken as a method of obtaining personality information from physical information. That is, the weight information 110 is weighed by the mat type weighing scale 160 as a measure of weight, and the seat position signal 120 is weighed by the seat detector 170 as a measure of height. Then, using the correlation between the weight and height and the voice, the voice feature conversion unit 330 converts the voice into personality information. The personality information includes a distribution of acoustic models used for recognition (a distribution of characteristic parameter values), or simply pitch information that defines voice pitch information. Thereafter, as described above, the merging is performed in the merging process 340 with the personality information obtained from the voice, and the acoustic model adapting unit 350 adapts the acoustic model. Further, the speech recognition processing section 500 executes a speech recognition process, and outputs a recognition result 999. When performing acoustic model constraints (selection), conversion and speech recognition,
The acoustic model 400 is read, the acoustic model is modified in the acoustic model conversion, and the acoustic model is used in the speech recognition.

FIG. 5 shows an acoustic model adaptation unit 3 according to the present invention.
It is a block diagram showing an example of 50. Merging is performed in the merging process 340, and the personality information is input, and the voice feature comparison unit 3510 compares the personality information with the standard acoustic models 400 already stored. Of the acoustic model adaptation execution unit 3530, such as selection and conversion of the acoustic model, is determined by the adaptation determination unit 3520. In the present embodiment, as the adaptive processing, for example, there is an acoustic model restriction unit 3531 or an acoustic model conversion unit 3532. Acoustic model constraint unit 3531
In, a set of acoustic models that are similar to the features of the voice obtained from the user's personality information is selected and set from a plurality of stored acoustic models. In the acoustic model conversion unit 3532, based on the characteristics of the user's voice,
The existing acoustic model is converted and set as a new acoustic model. The newly set acoustic model is newly stored in the acoustic model storage unit 400 and used in the next speech recognition processing unit.

The use of the acoustic model restriction unit 3531 and the acoustic model conversion unit 3532 is performed, for example, as follows. That is, in a device having a sufficient memory size to store a plurality of acoustic models, a large number of acoustic models corresponding to a large number of users are recorded, and the acoustic model constraint unit matches the actual user. An acoustic model is determined. In a device without sufficient recording memory, a recorded acoustic model is converted to calculate an acoustic model that matches the user's voice.

FIG. 6 is a block diagram showing an example of the speech recognition processing section 500. In the present embodiment, a continuous Hidden Markov Model (Hi
ddenMarkov Models). The continuous Hidden Markov Model is described in detail in the document "Speech Recognition by Probabilistic Model" (by Seiichi Nakagawa, edited by The Institute of Electronics, Information and Communication Engineers). As described with reference to FIG. 4, a Hidden Markov Model is used by using an acoustic model 400 adapted to a speaker according to the present embodiment, using a speech feature pattern obtained by sampling an input speech signal as an input. Is performed. That is, the probability distribution calculation unit 510 calculates the distribution probability of the acoustic model for the feature pattern of the input voice. Further, the probability accumulation unit 520 accumulates the probability distribution of the acoustic model corresponding to the series of words described in the word dictionary 600, and determines the cumulative probability of each entry in the word dictionary. Then, the judgment unit 530 outputs the word entry with the highest probability as the recognition result 999. In the present embodiment, word recognition has been described as an example, but for example, phrase recognition and sentence recognition are also achieved by using a word dictionary as a character sequence of phrases and sentences.
This is realized in the same way as word recognition.

Once speaker adaptation has been performed, there is basically no need to perform speaker adaptation unless the speaker changes, so there is no need to perform processing for extracting personality information and applying acoustic models. is there.

FIG. 7 is a diagram showing an example of the height / weight etc. detecting section 320 and the voice feature converting section 330 according to the present embodiment. The weight signal (weight value) 110 and the seat position signal 120 are generated by the mat type weight scale 160 and the seat position detector 170 of the height / weight etc. detecting unit 320, respectively. The seat position signal 120 is converted into height value information 130 by a height conversion unit 180. After that, these two pieces of information are converted into a voice feature pattern by the voice feature conversion unit 330. Specifically, the voice feature comparison unit 3310 uses the voice feature correspondence data 3320 describing the correspondence between the voice feature and the weight / height value, and determines whether the voice feature most corresponding to the weight value and the height value is gender. The voice feature parameters are separately extracted, and the voice feature conversion parameter calculating unit 3330 obtains the voice feature parameters correspondingly.

As the voice feature correspondence data 3320, for example, as shown in FIG. 12, a table in which a fundamental frequency (pitch) range as an acoustic feature corresponding to each of weight and height for each gender is provided in advance. Have been.

The speech feature parameters are, for example, feature parameters corresponding to an acoustic model such as a specific syllable or phoneme. Based on this information, the distribution of the entire acoustic model is adapted to the speech model of the speech recognition user. Will be adapted. In this case, speech recognition is performed using different acoustic models for male and female voices, and the better result is selected.

FIG. 8 is a diagram showing the concept when this embodiment is applied to car navigation. The car navigation device 5000 includes a display unit 5010, a speaker 502
0, an earphone jack 5030, a selector 5040, and a microphone 5050.
The concept of speaker adaptation is to adapt an existing acoustic model based on the voice of the user of the car navigation device input from the microphone 5050, and to detect a weight mat or seat position provided in the driver's seat of the car or the like. The acoustic model is adapted based on the input of the meter. When adapting an existing acoustic model based on the user's voice, input the voice of a greeting such as "Good morning" or "Today is" and input the voice model so that it is close to the voice of the speaker. It is possible to adapt. The speaker 5020 and the earphone jack 5030 function as a device that outputs a response voice from the system. Further, the selector 5040 is provided for setting whether or not to execute the speaker adaptation means prior to the speaker's speech recognition.

FIG. 9 is a diagram showing an example of a speech recognition performance evaluation result indicating speaker dependence in speech recognition. Recognition rate is not the same value for all speakers.
Due to habits such as voice accents, different performance values are shown depending on the speaker. In FIG. 9, the recognition rate is from 80% or less to 10%.
It can be seen that the distribution is close to 0%, and the result is that the number of speakers is the largest around 90%. The present embodiment can efficiently perform speaker adaptation of an existing acoustic model or the like, and can improve recognition performance even for a speaker having a low recognition rate.

FIG. 10 is a conceptual diagram showing an example of a speaker adaptation method used in the present embodiment. When the feature space 6100 of the existing HMM acoustic model has a distribution as shown in the figure, the adaptation according to the present embodiment converts and adapts to the acoustic model space 6200 of the specific speaker.
In the present embodiment, the speech unit of the acoustic model is a phoneme piece 6000 shorter than a phoneme. In the present embodiment, speaker adaptation of an acoustic model expressing such a phoneme 6000 or the like is performed.

FIG. 11 shows the result of speaker adaptation.
It is a conceptual diagram showing an example of improvement of recognition performance. In this example, an example is described in which the acoustic model is speaker-adapted on the condition that a predetermined word is uttered. The horizontal axis indicates the number of adaptive words, and it can be seen that as the number of adaptive words increases, the average recognition rate and the recognition rate for the lowest speaker having the lowest recognition rate also increase.
In particular, when the number of adaptive words further increases, the recognition rate of the lowest speaker approaches the average recognition rate, indicating that the effect of speaker adaptation is large. This example is an evaluation result obtained as a result of actual speaker adaptation. In the present invention, the recognition rate can be further improved by performing speaker adaptation in consideration of personality information such as weight and height.

FIG. 4 illustrates an example in which pitch information is extracted by the personality information extraction unit 310. In addition to the pitch information,
Alternatively, gender may be extracted instead of pitch information. This modification will be described with reference to FIGS.

FIGS. 13A and 13B show vowels / a
It is a graph which shows the frequency analysis result which represented the formant of / and / i / typically. A formant is a peak of a resonance frequency spectrum on a frequency axis. In FIG. 13, the first formant (F1), the second formant (F2), and the third formant (F3) are set from the lower frequency side. The formants and the shape and inclination of the entire spectrum differ depending on gender, and also differ between individuals.

FIG. 14 shows the first formant on the horizontal axis and the second formant on the second axis.
FIG. 3 shows a schematic diagram of a distribution of five vowels on an F1-F2 plane with a formant on a vertical axis. The F1-F2 plot points of each vowel differ from person to person. Also, in general, the plot points for each vowel are included within a range generally indicated by a rectangular area in FIG. In general, it is known that an adult male voice has both lower F1 and F2 than an adult female voice. From this, the personality information extraction unit 3 in FIG.
At 10, it can be inferred whether the input voice is male or female. Therefore, in this example, the acoustic model adaptation unit 350
, A male or female acoustic model can be selected based on the determination result of the input voice.

Although the present invention has been described with reference to an example in which the present invention is applied to a car navigation system, the present invention is not limited to this. For example, the present invention can be applied to a public service terminal. (Here, “public service” is not intended to exclude services provided by private companies, but means services provided publicly.) FIG. 15 shows a configuration example of such a public service terminal 1500. This public service terminal 1500 is used by the user 15
Numeral 10 is for receiving a specific service on the screen of the display unit 1503 by operating a touch panel, operation buttons, or the like (not shown), and a mat type weight scale 1505 and a user 1510 are provided on the floor where the user is located. A video camera 1502 is arranged at the upper front. User 15
A microphone 1504 for collecting ten voices is also provided. These elements are centrally managed under the processing device 1501, and weight information, image information, and audio information are captured. The processing device 1501 obtains approximate user height information from the image information. Therefore, speaker adaptation can be performed based on the obtained weight information and height information in the same manner as in the above example.

Although the preferred embodiment of the present invention has been described above, various modifications and changes can be made without departing from the gist of the present invention. For example, in the application example to the car navigation apparatus described above, the voice recognition of the driver's utterance was performed, but the function of detecting the weight and the seat position is also provided in the passenger seat, so that the driver and the passenger can be recognized. A means may be provided for the user to switch between them depending on which utters.

[0038]

According to the present invention, the speaker adaptation can be performed more reliably, the performance of the recognition rate can be improved, and as a result, a speech recognition application system with good operability can be provided.

[Brief description of the drawings]

FIG. 1 is a diagram showing the concept of conventional speaker adaptation that led to the invention.

FIG. 2 is a conceptual diagram of a process showing a speaker adaptation method according to the present invention.

FIG. 3 is a diagram showing an example of means for detecting a physical characteristic of a speaker, which is a main point of the present invention.

FIG. 4 is a block diagram showing an example of the configuration of the present invention in detail.

FIG. 5 is a block diagram showing an example of an acoustic model adaptation unit shown in FIG. 4;

FIG. 6 is a block diagram illustrating an example of a speech recognition processing unit illustrated in FIG. 4;

FIG. 7 is a diagram showing an example of a height / weight detection unit and a voice feature conversion unit according to the present invention.

FIG. 8 is a diagram showing the concept when the present invention is applied to car navigation.

FIG. 9 is a diagram showing an example of a speech recognition performance evaluation result indicating speaker dependence in speech recognition.

FIG. 10 is a conceptual diagram showing an example of a speaker adaptation method used in the present invention.

FIG. 11 is a conceptual diagram showing an example of improvement in recognition performance.

FIG. 12 is an explanatory diagram showing an example of audio feature correspondence data according to the present invention.

FIG. 13 is an explanatory diagram of a formant configuration of vowels / a / and / i /.

FIG. 14 is an explanatory diagram showing F1-F2 distribution of vowels of men and women.

FIG. 15 is a diagram showing a configuration when the present invention is applied to a public service terminal.

[Explanation of symbols]

100: voice signal, 110: weight signal, 120: seat position signal, 150: seat, 160: mat-type weight scale, 170: seat position detector, 180: height conversion unit,
200: voice input unit, 210: voice analysis unit, 300: acoustic model adaptation process, 310: voice personality information extraction process, 320: physical feature detection unit, 330: voice feature conversion unit, 3310: voice feature comparison Unit, 3320: voice feature correspondence data, 3330: voice feature conversion parameter calculation unit, personality information fusion unit: 340, 350: acoustic model adaptation unit, 3510: acoustic feature comparison unit, 3520: adaptation determination unit, 3530: acoustic model Adaptive execution unit, 3531 ... Acoustic model constraint unit, 3532 ... Acoustic model conversion unit, 400 ...
Acoustic model, 500: voice recognition processing, 510: probability distribution calculation unit, 520: probability accumulation unit, 530: determination unit, 600
... word dictionary, 999 ... recognition result, 2000 ... voice recognition system with speaker adaptation function, 5000 ... car navigation device, 5010 ... display, 5020 speaker, 503
0: Earphone jack, 5040: Selector, 5050
... microphone, 6000 ... phoneme, 6100 ... HMM
Model feature space, 6200: Feature space of specific speaker model.

 ──────────────────────────────────────────────────の Continuing on the front page (72) Inventor Akio Amano 1-280 Higashi Koigakubo, Kokubunji City, Tokyo Inside the Central Research Laboratory of Hitachi, Ltd. (72) Inventor Shinya Oba 5--20-1, Josuihoncho, Kodaira-shi, Tokyo Semiconductor Division, Hitachi, Ltd.

Claims (10)

[Claims] 1. A speaker adaptation method in a speech recognition system equipped with a speech recognition function, comprising detecting physical characteristic information of a speaker, and characterizing the speaker's voice in accordance with the physical characteristic information. A speaker adaptation method in a speech recognition system, comprising guessing and performing speech recognition using an acoustic model of the speech recognition system suitable for the estimated speech feature. 2. The method according to claim 1, wherein the physical characteristic information is at least one of weight information and height information of the speaker. 3. A speech feature of a speaker is identified from the utterance content of an adaptation word specified in advance by a speaker, and the speech feature obtained based on the speech feature and the physical feature information of the speaker is identified. 3. The selection or conversion of the acoustic model based on a feature of a voice combined with a feature.
A speaker adaptation method in the described speech recognition system. 4. A speech recognition system equipped with a speech recognition function, comprising: means for acquiring physical characteristic information of a speaker; a correspondence table in which the physical characteristic information of the speaker is associated with voice characteristics; Means for extracting the features of the speaker's voice by illuminating the acquired physical feature information with the correspondence table; a plurality of acoustic models for speech recognition; and an acoustic model adapted to the extracted features of the voice. Acoustic model adaptation means for selecting or converting from among the plurality of acoustic models, and speech recognition processing means for performing speaker speech recognition processing using the acoustic model obtained by the acoustic model adaptation means. A speech recognition system characterized by the following. 5. The speech recognition system according to claim 4, wherein said physical characteristic information is at least one of weight information and height information of a speaker. 6. The voice recognition system is mounted on a car navigation device, the means for acquiring weight information includes a mat-type weighing scale disposed on a speaker's seat, and the means for acquiring height information includes a speech. 6. The voice recognition system according to claim 5, further comprising a seat position detector for detecting a position of a seat of a person. 7. The means for obtaining weight information includes a mat-type weight scale arranged at a place where a speaker is located, and the means for obtaining height information includes a video camera for obtaining an image of a speaker and the video camera. 6. The speech recognition system according to claim 5, further comprising processing means for obtaining height information by processing an image. 8. A speech feature of a speaker is identified from the utterance content of an adaptation word specified in advance by the speaker, and the speech feature and the speech obtained based on the physical feature information of the speaker are identified. 5. A speech recognition system according to claim 4, wherein said speech model adaptation means selects or converts said speech model based on said combined speech features. 9. The speech recognition system according to claim 4, wherein at least two types of said acoustic models are provided for men and women. 10. The apparatus according to claim 4, further comprising a selection means for allowing a user to select whether or not to apply the speaker adaptation of the acoustic model before the speech recognition of the speaker. Voice recognition system.