JP2000206988A - Voice processing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice processing device having a speaker recognition function that surely specifies a non-registered speaker who enters into voice data. SOLUTION: Feature parameters of voice for registration are stored in a voice model recording section 13B as registered voice models for each speaker. A feature parameter extracting section 14 extracts feature parameters from the voice data recorded in a voice data recording section 13A. Then, the degreee of similarity between the feature parameters of the extracted voice data and the registered voice model recorded in the section 13B is obtained by a degree of similarity computing section 16. Based on the computational result in the section 16, a speaker specifying section 17 discriminates whether he is a registered speaker corresponding to the registered voice model or a non-registered speaker.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech processing apparatus having a speaker recognition function.

[0002]

2. Description of the Related Art Conventionally, as an audio processing device, an audio signal input from a microphone or the like is converted into a digital signal, and the digital signal is recorded on a recording medium such as a semiconductor memory. 2. Description of the Related Art A digital audio recording / reproducing apparatus which converts a signal into a signal and outputs the signal is known.

In such an audio processing apparatus, it is desired to improve operability and searchability when reproducing recorded audio data, and various proposals have been made. As disclosed in Japanese Patent Application Laid-Open No. 10-3300, index information for designating a range to be reproduced is recorded in a semiconductor memory, and when an operation such as fast forward is performed, only the designated range is determined based on the recorded index. There is a known device for playing back.

However, in such a configuration, a portion to be reproduced must be specified from the recording range, so that there is a problem that the operability of reproduction is extremely poor.

[0005] Recently, with the development of speech processing technology, speech recognition technology and speaker recognition technology have been put into practical use.
As an application of these recognition techniques, for example, as disclosed in Japanese Patent Application Laid-Open No. 8-153118, a speaker recognition function can be used to search for voice data of a specific speaker and reproduce only voice data of the corresponding speaker. By doing so, it is considered that the operability and searchability during reproduction are enhanced.

[0006]

However, in a speaker having such a speaker recognition function, a designated speaker is selected from pre-registered speakers and reproduced.
For example, if unregistered speakers appear during a conversation and these unregistered speakers are recorded together with registered speakers,
If a specific speaker to be played is specified, the speaker with the characteristics closest to the specific speaker may be selected and played back from among the unregistered speakers, and recorded contents other than the specified speaker will also be played. There was a problem that it would.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and has as its object to provide a speech processing apparatus having a speaker recognition function capable of reliably specifying an unregistered speaker appearing in speech data. And

[0008]

According to the first aspect of the present invention,
Voice input means for inputting voice, registered voice model recording means for recording, for each speaker, feature parameters of the registered voice as a registered voice model, and extracting feature parameters from the voice signal input from the voice input means And a similarity parameter for obtaining a similarity between a feature parameter of a voice signal input from the voice input means extracted by the feature parameter extracting means and a registered voice model recorded in the registered voice model recording means. And a speaker judging unit for judging a registered speaker or an unregistered speaker corresponding to the registered voice model from the calculation result by the similarity calculating unit.

According to a second aspect of the present invention, there is provided voice input means for inputting voice, registered voice model recording means for recording characteristic parameters of voice for registration as a registered voice model for each speaker, and said voice input means. Voice data recording means for recording voice data input from the voice data, feature parameter extraction means for extracting feature parameters from voice data recorded in the voice data recording means, and the voice data extracted by the feature parameter extraction means. A similarity calculating means for calculating a similarity between the characteristic parameter of the voice data recorded in the recording means and the registered voice model recorded in the registered voice model recording means; and the registered voice based on the calculation result by the similarity calculating means. It comprises speaker determination means for determining a registered speaker or an unregistered speaker corresponding to the model.

[0010] The third aspect of the present invention is the first or second aspect.
In the invention described, the speaker determination means obtains the maximum value of the similarities between the registered voice models of the speakers by the similarity calculation means, and if the maximum value is equal to or greater than a predetermined threshold,
The current input voice is determined to be a registered speaker corresponding to the registered voice model, and if the maximum value is less than a predetermined threshold, the speaker of the current input voice is determined to be an unregistered speaker. I have.

As a result, according to the present invention, even when unregistered speakers are present together with registered speakers in voice data, unregistered speakers appearing in voice data are identified and distinguished from registered speakers. can do.

[0012]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows a schematic configuration of a digital audio recording / reproducing apparatus to which the present invention is applied. In the figure, 1
Is a microphone as an audio input means, and the microphone 1 is composed of a preamplifier 2, a low-pass filter 3A
It is connected to the encoding / decoding unit 7 via the / D converter 3B. Here, the encoding / decoding unit 7 performs, for example, CEL
P (Code Excited Linear Pre
Encoding / decoding of speech is performed by a digital coding (DSP) method, and a DSP (Digital Signa) is used.
1 Processor). The encoding / decoding unit 7 includes a D / A converter 3 of the CODEC 3.
The speaker 5 is connected via C, a low-pass filter 3D, and a power amplifier 6.

A system controller 10 is connected to the CODEC 3. The system control unit 10 controls the operation of each unit, and is realized by, for example, an 8-bit CPU.

The encoding / decoding unit 7 and the flash memory 13 are connected to the system control unit 10 via a memory control unit 8 realized by, for example, a 16-bit CPU, and the CODEC 3 and the display unit 9 are connected to the system control unit 10. , An operation input unit 11 and a power supply 12 are connected.

Here, the memory controller 8 controls the input and output of signals between the flash memory 13 and the encoder / decoder 7. The flash memory 13 has three areas of a voice data recording unit 13A, a voice model recording unit 13B, and a speaker code recording unit 13C, and together with encoded data header information output from the memory control unit 8,
The audio data is recorded in the audio data recording unit 13A. The flash memory 13 may be built in the digital audio recording / reproducing device or may be detachable. The operation input unit 11 includes operation buttons or operation switches for recording, reproducing, speaker registration, and the like.

The encoding / decoding section 7 is connected to a feature parameter extracting section 14. This feature parameter extraction unit 14 uses an expression format suitable for speaker recognition from voice data,
For example, it extracts feature parameters such as pitch and cepstrum.

A speech model creation unit 15 is connected to the feature parameter extraction unit 14. The speech model creation unit 15 receives the time-series data of the feature parameters from the feature parameter extraction unit 14, creates a standard pattern of the feature parameters as a registered speech model, and flashes this speech model via the memory control unit 8. The voice model is recorded in the voice model recording unit 13B of the memory 13.

The feature parameter extracting unit 14 is connected to a similarity calculating unit 16. This similarity calculator 1
6 includes a voice model recording unit 13 of the flash memory 13.
B is connected. The similarity calculating unit 16 calculates the similarity between the voice model of each speaker recorded in the voice model recording unit 13B of the flash memory 13 and the feature parameter.

The similarity calculation section 16 is connected to a speaker identification section 17 as speaker determination means. The speaker identification unit 17 identifies a registered speaker or an unregistered speaker from the calculation result of the similarity calculation unit 16, and the speaker code corresponding to the registered speaker or an unregistered speaker representing the unregistered speaker. The code is recorded in the speaker code recording unit 13C of the flash memory 13 via the memory control unit 8.

Next, the operation of the embodiment configured as described above will be described.

First, when an operator operates a recording operation button (not shown) of the operation input unit 11, an audio analog signal input from the microphone 1 is amplified by the preamplifier 2, and the audio signal component is amplified by the low-pass filter 3A. Unnecessary high frequency components are removed, and are converted to digital signals by the A / D converter 3B. Then, in the encoding / decoding unit 7,
The digital signal from the A / D converter 3B is encoded, and the encoded data is recorded in a predetermined area of the audio data recording unit 13A of the flash memory 13 via the memory control unit 8. . In this case, header information is recorded in the audio data recording unit 13A together with the encoded data.

Next, when the operator operates a reproduction operation button (not shown) of the operation input unit 11, the memory control unit 8
Audio data recording unit 13 of flash memory 13
A. The encoded data is read from A, and the encoding / decoding unit 7
Thus, decrypted data is generated. This decrypted data is
The D / A converter 3C converts the signal into an analog signal. The low-pass filter 3D removes unnecessary high-frequency components from the frequency components included in the analog signal. The signal is amplified by the power amplifier 6 and output as a reproduction signal from the speaker 5. Is done.

The above are the recording and playback operations, which are the basic operations of the digital audio recording and playback device. Next, the speaker registration process used for speaker recognition processing, which is a feature of the present invention, will be described. The operation will be described. Note that this speaker registration is performed for each speaker to be recorded prior to the above-described recording operation.

First, speaker registration will be described.

In this case, when the operator operates a speaker registration operation button (not shown) of the operation input section 11, the flowchart shown in FIG. 2 is executed.

In step 201, the speaker registration voice data is input to the feature parameter extraction unit 14. The speaker registration voice data at this time may be input by the speaker from the microphone 1, or the speaker registration voice data may be recorded in advance, and the recorded data may be input. .

Next, at step 202, the feature parameter extracting unit 14 extracts the input registration voice data into an expression format suitable for speaker recognition, for example, a feature parameter such as pitch or cepstrum.

Then, the process proceeds to step 203, where the time-series data of the characteristic parameters is input to the speech model creating unit 15, and a standard pattern of the feature parameters is created as a speech model. 8 through the audio model recording unit 13B of the flash memory 13.

Next, the voice model recording unit 1
A process for recognizing a speaker with respect to voice data recorded in the voice data recording unit 13A of the flash memory 13 using the voice model of each speaker stored in 3B will be described.

In this case, when speaker recognition processing is instructed,
The flowchart shown in FIG. 3 is executed.

First, in step 301, the value of a counter variable f of a frame of audio data is set to zero. next,
In step 302, the value of f is incremented by one, and in step 303
Then, it is determined whether or not the sound of the frame corresponding to f is sound. The determination method here may be, for example, a method disclosed in Japanese Patent Application Laid-Open No. 9-281987,
As disclosed in JP-A-9-218694, sound / non-sound information recorded in advance in a header corresponding to a frame of audio data may be used.

If it is determined in step 303 that no audio data is recorded in the current frame and that there is no sound, no processing is performed on the current frame, and the flow returns to step 302 to return to the value of f. , And the operation after step 303 is performed for the next frame.

On the other hand, if it is determined in step 303 that audio data is recorded in the current frame, step 30
4, the address information of the audio data recording unit 13A in which the audio data of the current frame is recorded is stored in the flash memory 1
3 is recorded in the speaker code recording unit 13C.

Next, in step 305, the speech data of the current frame is input to the feature parameter extraction unit 14 to extract predetermined feature parameters. In step 306, the similarity calculation unit 16 is used to extract the predetermined feature parameters in step 305. The similarity between the feature parameter obtained and the voice model of each speaker recorded in the voice model recording unit 13B of the flash memory 13 is calculated. Then, in step 307, the speaker specifying unit 17
In step 308, it is determined whether or not the similarity of the maximum value is equal to or greater than a predetermined threshold.

If it is determined that the maximum value is equal to or larger than the predetermined threshold value, in step 309, the speaker of the audio data of the current frame determines the similarity of the maximum value among the audio models recorded in the audio model recording unit 13B. The speaker corresponding to the voice model having the degree is specified. Then, in step 310, the speaker code corresponding to the speaker is recorded in the speaker code recording unit 13C of the flash memory 13 so as to correspond to the address information of the audio data of the current frame recorded in step 304. Next, in step 311, it is determined whether or not the audio file has ended. If the audio file has not ended, the process returns to step 302, increments the value of f by 1, and performs the operations of step 303 and subsequent steps on the next frame. If the file is over, the process ends.

On the other hand, if it is determined in step 308 that the maximum value is less than the predetermined threshold value, the flow advances to step 312 to specify that the speaker of the voice data of the current frame is not a registered speaker. Then, in step 310, the unregistered speaker code corresponding to the speaker is recorded in the speaker code recording unit 13C of the flash memory 13 so as to correspond to the address information of the audio data of the current frame recorded in step 304. Next, in step 311, it is determined whether or not the audio file has ended. If the audio file has not ended, the process returns to step 302 to add 1 to the value of f, and performs the operations of step 303 and subsequent steps for the next frame. If the file is over, the process ends.

Accordingly, even when the registered speaker and the unregistered speaker are mixed in the voice data, the unregistered speaker appearing in the voice data is specified to be distinguished from the registered speaker. Since the unregistered speakers of the related art are recorded together with the registered speakers, and the specific speaker to be reproduced is specified, the characteristic of the unregistered speaker closest to that of the specific speaker can be obtained. When a specific speaker is specified, only the specified speaker can be accurately reproduced, compared to a case where a selected speaker is also selected and recorded contents other than the specified speaker are also reproduced.

Although the speaker is recognized based on the voice data recorded in the voice data recording unit 13A of the flash memory 13, the amount of calculation required for the speaker recognition process is generally enormous. Inexpensive devices can be very advantageous.

Further, since the speaker is recognized with respect to the voice data recorded in the voice data recording unit 13A, it is possible to give a degree of freedom to the execution timing of the speaker recognition processing. It can be performed automatically immediately after the process is completed, or can be performed by a speaker recognition operation of the operation input unit 11 by the operator.

Further, the speaker recognition processing is performed only for the sound section, but the silent section does not have the individuality of the speaker, and the calculation of the similarity is unnecessary. Since the speaker recognition processing is performed, accurate speaker recognition can be realized.

In the above-described embodiment, the speaker is recognized with respect to the audio data recorded in the audio data recording unit 13A of the flash memory 13.
It goes without saying that the speaker can be recognized for the voice data directly input from the microphone 1.

[0043]

As described above, according to the present invention, even when unregistered speakers are present together with registered speakers in voice data, unregistered speakers appearing in voice data are distinguished from registered speakers. Since the specified speaker can be specified separately, only the specified speaker can be accurately reproduced when the specified speaker is specified.

[Brief description of the drawings]

FIG. 1 is a diagram showing a schematic configuration of an embodiment of the present invention.

FIG. 2 is a flowchart for explaining the operation of the embodiment;

FIG. 3 is a flowchart for explaining the operation of the embodiment;

[Explanation of symbols]

 1. Microphone 2. Preamplifier 3. CODEC 3A. 3D low-pass filter 3B A / D converter 3C D / A converter 3D low-pass filter 5 speaker 6 power amplifier 7 decoding unit 8 memory control unit 9 display unit 10 system control unit 11 Operation input unit 12 Power supply 13 Flash memory 13A Voice data recording unit 13B Voice model recording unit 13C Speaker code recording unit 14 Feature parameter extraction unit 15 Voice model creation unit 16 Similarity calculation unit 17 Talk Person identification department

Claims (3)

[Claims] 1. A voice input means for inputting voice, a registered voice model recording means for recording a characteristic parameter of a voice for registration as a registered voice model for each speaker, and a voice signal input from the voice input means A feature parameter extracting means for extracting a feature parameter from the speech signal; a feature parameter of a speech signal input from the speech input means extracted by the feature parameter extracting means; and a registered speech model recorded in the registered speech model recording means. Similarity calculating means for obtaining a similarity; and speaker determining means for determining a registered speaker or an unregistered speaker corresponding to the registered voice model from a calculation result by the similarity calculating means. Audio processing device. 2. Voice input means for inputting voice, registered voice model recording means for recording a feature parameter of a voice for registration as a registered voice model for each speaker, and voice data input from the voice input means Audio data recording means for recording the audio data; characteristic parameter extraction means for extracting characteristic parameters from audio data recorded in the audio data recording means; and audio data recording means extracted by the characteristic parameter extraction means. A similarity calculating means for calculating a similarity between the feature parameter of the voice data and the registered voice model recorded in the registered voice model recording means; and a registration story corresponding to the registered voice model based on the calculation result by the similarity calculating means And a speaker determination unit for determining a speaker or an unregistered speaker. 3. The speaker determination means obtains the maximum value of the similarity between each speaker and the registered voice model by the similarity calculation means, and if the maximum value is equal to or more than a predetermined threshold, the current input value is determined. The voice is determined to be a registered speaker corresponding to the registered voice model, and if the maximum value is less than a predetermined threshold, the speaker of the current input voice is determined to be an unregistered speaker. Item 3. The audio processing device according to item 1 or 2.