JPS59111699A - Speaker recognition method - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field of the Invention The present invention relates to a speaker recognition system, and in particular, when the voices pronounced by a plurality of people are registered, the input voice is recognized as one of the registered voices of the plurality of people. It relates to something that can be recognized to be similar to any one of them.

Background of the Technology Current speech recognition devices can achieve a high recognition rate when nine speakers use a dictionary in which their own voices are registered, but they cannot recognize anyone's voice, and are unable to recognize anyone else's voice. If you use a different dictionary, the recognition rate will decrease. Therefore,9 For example, when performing speech recognition using a voice transmitted over the telephone, the first step is to recognize who the speaker is on the telephone, or to which of multiple registered speakers the speaker is similar. 9 speaker recognition was required.

Conventional techniques and problems Conventional speaker recognition methods include methods that focus on "speech style" and methods that focus on "speech style";
There is a way to focus on sound quality. The former.

For example, there is a method that focuses on patterns of changes in speech rate and in)/nation, but although it is a simple method, registered dictionaries contain a lot of data related to sound quality, and data on "speech style" includes sound quality. Because there is insufficient analysis regarding this, it is not suitable for selecting a dictionary prior to using a speech recognition device. The latter is a method that focuses on the sound quality determined by the shape of the speaker's vocal cords and the shape of the resonator tube such as the oral cavity. Although it is unsuitable in terms of accuracy for determining whether
This is suitable for determining whose voice the voice of a new speaker who has not been registered resembles.

Conventional speaker recognition technology that focuses on sound quality analyzes the sound quality of two utterances at each frame interval, extracts % total parameters, and then adds up and averages these characteristic parameters along the time axis. There is a way to compare each person. However, with this method, the averaged pattern contained considerable percentage characteristics of speaking style, such as the influence of vocalization method, such as the length and shortness of the sun, in addition to sound quality characteristics, and was insufficient for accurate recognition.

Purpose of the Invention The purpose of the present invention is to solve this problem by analyzing the voice of a registered speaker at each frame interval, storing the obtained parameters in memory, and registering them for each frame of the input voice. By determining whose voice the speaker's voice resembles, and making a comprehensive judgment based on which frame resembles the voice of the most people after the input voice is uttered, the system is not affected by the vocalization method.
It is an object of the present invention to provide a speaker recognition method that can highly accurately determine to whom a speaker's voice is similar.

Structure of the Invention In order to accomplish this purpose, the speaker recognition method of the present invention has two steps:
In a speaker recognition device that processes the voice pronounced by Iruma and determines which of the registrants the voice of the two speakers resembles,
A subclassified sound type pattern memory is provided to store a parameter time series for each speaker, which is obtained by analyzing the voices uttered by multiple people at frame intervals.

A selection means is provided for calculating the correlation between the parameter time series obtained by analyzing the speaker's voice in frame cycles and the parameter time series of the subclassified sound pattern memory, and selecting a registered speaker name with a large correlation in KM for each frame cycle. The present invention is characterized in that the name of the registered speaker selected most times after nine utterances is determined and this is output as the speaker recognition result.

Summary of the Invention In the present invention, a group of patterns obtained by analyzing voices uttered by a plurality of speakers for each frame interval is analyzed in advance for each speaker.
C are organized and stored in a memory (referred to as subclassification note type memory). Then, the correlation (degree of similarity) is calculated between the analysis result for one frame of the speaker to be recognized and all the items in the subclassified sound type memory.

Record the vocalization name of the pattern with the highest degree of similarity.

Such processing is performed on all frames of the voice of the speaker to be recognized, and the name of the speaker selected most frequently is used as the speaker recognition result.

History of invention & April An embodiment of the present invention will be described in detail based on the accompanying drawings.

In the figure, 1 is a microphone, 2 is a 16-channel band-pass filter bank (hereinafter referred to as a band-pass filter), and 3 is a multiplexer.

4 is an analog-to-digital converter (hereinafter referred to as an A/D converter), 5 is a subclassification note type memory, 6 is a Chebyshev norm calculation circuit, 7 is a minimum value calculation unit, 8 is a decoder, 9 is a registered speaker frequency record part, ioI'i maximum value calculation part 181.8
2 are switch parts.

The bandpass filter 2 analyzes the audio signal input from the microphone 1 into 16 frequencies from f□ to f1g, and the bandpass filter 2 analyzes the audio signal input from the microphone 1 into 16 frequencies from f□ to f1g.
It outputs 4 analog signals.

The multiplexer 3 performs time-division sampling by scanning the analog signals of the channels 1t-1 to 16, for example, once every 10 m5 sampling period. This time-divided one analog signal output is converted into a digital quantity by the A/D converter 4, and is output as a digital quantity of, for example, 16 words/frame. Therefore, if the input utterance length is, for example, 1 second, then 1 o for 1 utterance.
o frames x 16 words - 1600 words will be output.

The subclassified note type memory 5 is a memory that holds the characteristics of registrants, and is prepared for each registrant in order to hold the characteristics of each registrant. Therefore, when there are 10 registrants as in this example, the first memory 5-0 for subclassified sound m to the tenth memory 5-9 for subclassified sound m are prepared.

The Chebyshev norm calculation circuit 6 calculates the similarity of frames and divides ΣlI+−D+1 −1. Here, I represents the first channel output from the A/D converter 4, and Dl represents the first channel of one dictionary item held in the subclassification note type memory 5 transmitted via the switch section S2. Indicates word. This calculation result is one frame data (16 words) of the recognized speech sent from the A/D converter 4 and one dictionary item (16 words) sent from the subclassification sound type memory 5 via the switch section S2. The distance is expressed as K. Once every 10 msec, when one frame worth of data is transmitted from the A/D converter 4, the Chebyshev norm calculation plane circuit 6 switches the switch section S2 to subdivide note type first memories 5-0 to 5-0. Classification sound type 10th memory 5
-9 side, 100 items x 10 (m-iooo
The distance calculation for the item is performed, and the minimum value calculation unit 7 calculates the minimum value of the 1000 calculation results, and the data giving the minimum value is stored in the subclass note type first memory 5-0 to subclass note type No. A 4-bit identification code indicating from which of the 10 memories 5-9 the frame is output is output for each frame. In other words, the minimum value calculation section 7 calculates 4 times once every 10 msec.
The bit identification code will be output.

The decoder 8 deciphered this 4-bit identification code, and when it compared it with, for example, the data output from the subclassified note type No. 17 mori 5-0, it determined that the minimum value was assigned. At this time, the output is sent to the first counter 9-0 of the registered speaker frequency recording unit 9, and it is incremented by 1.9 When the output is interpreted as being output from the second memory 5-1 for subclassified note types, the second counter 9 Send output to -1. In this way, the first counter 9-0 to the tenth counter 9-9 count at least similar registered speakers for each frame, and the maximum value among these counters is calculated as the maximum value. By detecting it in the unit 10, it can be determined whether the recognized speech is extremely similar to any of the registered speakers.

The present invention will now be described in detail with reference to the accompanying drawings.

(1) At the time of registration At the time of registration, the switch section S1 is first connected to the subclassified sound type first memory 5-0, and the first registered speaker is made to emit, for example, a predetermined voice. This voice is input from microphone 1 and passed through bandpass filter 2.
The frequency is analyzed into 6 channels, and 16 channels of analog signals 3 are output. Time-division sampling is performed by scanning the analog signals of channels 1 to 16 once in a sampling period of 10n1 seconds using the multiplexer 3.

This output is converted into a digital quantity by an A/D converter 4. In this way, the A/D converter 4 generates nine digital outputs, one word for each channel every 10 msec, totaling 16 words, which are registered in the subclassified note type first memory 5-0. will be done. Therefore, if the input utterance length is 1 second, 100 frames x 16 words (1600 words) will be registered for one utterance. Next, when a second registered speaker registers, the switch section S1 is connected to the subclassified note type second memory 5-1 side and similar input processing is performed, so that the subclassified note type second memorandum IJ
5-1 holds the characteristics of the second registered speaker.

This is done for each registered speaker, so when there are 10 registered speakers, the 10th subclassification sound type memory 5-9
The characteristics of each registered speaker will be maintained until then.

(2) When recognizing whether the input speech is similar to any of the registered speakers, the switch section S1 is opened. What is the audio input at this time?

It is desirable that the voice is the same as that uttered by the registered speaker when registering the characteristics in the subclassification sound type memory 5. This voice to be recognized inputted from microphone 1 is as described above (1).
), the frequency is analyzed into 16 channels, and these are divided into 1
It is scanned at a sampling period of 0 msec and converted into a digital quantity, and the digital output of 16 words per frame is transmitted to the Chebyshev norm calculation circuit B. At this time, the switch section S2 connects to the subclassified note type first memory 5-0, reads out this memory word by word, and uses the Chebyshev norm calculation circuit 6 to read out the above! l l I, -D, l
-c-1 Perform the financial power expressed. In other words, the sum of the absolute values of the differences between the corresponding terms of each item of 16 words of data obtained from the speech to be recognized and the subclassified sound type memory is subtracted, and this calculation result is sent to the A/D converter. This represents the distance between one frame of data (16 words) sent from 4 and one dictionary item (16 words) sent from subclassification note type memory 5. Once every 10 msec, when one frame worth of data is transmitted from the A/I) converter 4, the Chebyshev norm calculation circuit 6 sends the switch section S2 to the subclassification note type first memory 5-0 to subclassification. The distance is calculated for 100 terms x 10 sets by sequentially switching to the 10th note type memory 5-9 side.
The minimum value calculation unit 7 performs the minimum value of the 1000 calculation results and determines whether the data giving the minimum value is from any of the subclass note type first memory 5-0 to subclass note type tenth memory 5-9. For example, a 4-bit identification code indicating whether the output has been output is output.

That is, the minimum value calculation section 7 outputs this identification code once every 10 msec, but this identification code is decoded by the decoder 8, and the corresponding first counter 9-0 to tenth counter 9-9 selectively output the identification code. The de/-der 8 outputs a signal that is incremented by +1. When one voice to be recognized has been uttered in this manner, the maximum value calculation section 10 calculates the registered speaker frequency recording section 9.
The first counter 9-0 to the tenth counter 9 forming the
-9 is compared, and the counter number whose value is also large is output as the speaker recognition result, and the first counter 9-0 to the tenth counter 9-9 are reset.

In the above explanation, an example was explained in which a bandpass filter with 16 channels was used, but of course the number of channels is not limited to this, and it is possible to use an appropriate n channel or use a digital filter bank. It is also possible to change the frame period to another suitable time instead of 10 msec. Of course, the number of registered speakers is not limited to ten, and any number of speakers can be selected.

Furthermore, the voice uttered during speaker recognition may be specific or may be different at the time of registration and at the time of recognition.

Effects of the Invention According to the present invention, it is possible to perform speaker recognition based on sound quality without being affected by vocalization methods such as having long and short final utterances, so that highly accurate speaker recognition can be performed. can. Therefore, this makes it possible to improve the speech recognition rate for unspecified speakers by using dictionaries of similar registered speakers. Furthermore, the present invention performs preprocessing on speakers input using a telephone to select similar registered speakers, and then performs highly accurate speech recognition by using the registered dictionary. I can do it.

[Brief explanation of drawings]

The accompanying drawing is a configuration diagram of an embodiment of the present invention. In the figure, 1 is a microphone, 2 is a band-pass filter bank, 3 is a multiplexer, 4 is an analog-to-digital converter, 5 is a subclassified note type memory, 6 is a Chebyshev norm calculation circuit, 7 is a minimum value calculation unit, 8 is a decoder, 9
10 is a registered speaker frequency recording unit, and 10 is a maximum value calculation unit TSL.
＋ ”'2 are switch parts respectively. Patent applicant: Fujitsu Limited, representative patent attorney Akira Yamatani Sakae

Claims (1)

[Claims] A speaker recognition device that processes speech produced by a human and determines which of the registrants the speaker's voice resembles,
A subclassification sound type pattern memory is provided to store a parameter time series for each speaker, which is obtained by analyzing the voices of multiple people at a frame period. A selection means is provided for calculating the correlation with the parameter time series of the pattern memory and selecting the registered speaker name with the highest correlation for each frame period. This speaker recognition method is characterized by determining the name of a registered speaker who has been selected many times after utterance and outputting this as the speaker recognition result.