JP4408205B2 - Speaker recognition device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speaker recognition device that recognizes a speaker by inputting voice.
[0002]
[Prior art]
In general, in speaker recognition, input speech is cut into frames of about 30 msec at cyst intervals of about 10 msec, and various individuality features obtained for each frame are extracted as feature vector sequences over the entire voiced sound section. The extracted feature vector sequence is evaluated by a comparison method such as the VQ distortion method and the HMM method, for example, to perform speaker recognition.
In addition, recently, in order to recognize the individuality of speech, it is known that not only voiced sound sections but also unvoiced sound sections are used (for example, see Non-Patent Document 1).
[0003]
[Non-Patent Document 1]
Tomoko Matsui, Sadaaki Furui, IEICE Technical Report SP90-26, "Text independent speaker recognition using sound source and vocal tract features", IEICE, June 1990, pp.55-62
[0004]
[Problems to be solved by the invention]
By the way, it is known that weighting due to the feature amount is performed when performing speaker recognition. However, when such weighting is applied to one using both a voiced sound segment and an unvoiced sound segment, a frame to be divided While cutting out frames with a fixed length, weighting is applied not only to voiced sound intervals but also to unvoiced sound intervals, but for voiced sound intervals, for example, processing such as weighting twice that of unvoiced sound intervals is performed. become.
However, if the frame length for extracting both the unvoiced sound section and the voiced sound section is fixed, the analysis in the voiced sound section becomes rough. The rough portion is dealt with by giving weights, but such processing cannot sufficiently recognize the speaker. In addition, when calculating the distance between the analysis result and a preset speaker model, weighting processing must be performed each time, which is troublesome.
The present invention provides a speaker recognition device that can perform speaker recognition with high accuracy and can easily calculate the distance from the speaker model.
[0005]
[Means for Solving the Problems]
According to the present invention, a voice input unit including a microphone, for example, for inputting voice, and a frame cutout that cuts voice waveform data input by the voice input unit into frames having a set frame width at a set shift interval. Means, a feature vector generating means for converting the speech waveform data of the frame extracted by the frame extracting means into a feature vector, a pitch detecting means for detecting the presence or absence of a vocal cord frequency in the frame extracted by the frame extracting means, and a speaker in advance A speaker model storage means storing a model; and a recognition means for performing speaker recognition from a distance between a feature vector generated from the feature vector generation means and a speaker model stored in the speaker model storage means, and frame extraction When the pitch detecting means detects the vocal cord frequency, the means sets the shift interval to T corresponding to the unvoiced sound section. Instead of ₀ , T ₁ (T ₁ <T ₀ ) that does not depend on the pitch frequency is set.
[0006]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing the overall configuration of the apparatus. 1 is a CPU (central processing unit) that constitutes a control unit main body, 2 is a ROM (read only memory) that stores program data and the like that the CPU 1 controls each part. 3 is a RAM (Random Access Memory) provided with a memory used by the CPU 1 for data processing and various buffer memories.
[0007]
Also, 4 is an I / O port, 5 is a hard disk device provided with speaker model storage means for storing a number of speaker models in advance, 6 is a keyboard controller for fetching key signals from the keyboard 7, and 8 is a display 9. This is a display controller that displays information such as text and images.
The CPU 1, ROM 2, RAM 3, I / O port 4, hard disk 5, and controllers 6 and 8 are electrically connected to each other via a bus line 10.
[0008]
Further, a microphone 11 for capturing sound is provided. The microphone 11 converts the captured sound into an electrical analog signal and outputs it, and supplies it to the low-pass filter 12 at the next stage. The low-pass filter 12 cuts and outputs a frequency equal to or higher than a predetermined frequency from the input analog signal, and inputs the cut signal to the A / D converter 13 at the next stage. The A / D converter 13 converts the input analog signal into a digital signal at a predetermined sampling frequency and the number of quantization bits, and supplies the digital signal to the I / O port 4 as voice waveform data. The microphone 11, the low-pass filter 12, and the A / D converter 13 constitute an audio input means.
[0009]
FIG. 2 is a functional block composed of a composite of the CPU 1, ROM 2, RAM 3, and hard disk device 5. The frame cutout means 21 shifts the audio waveform data input from the I / O port 4 with a set shift. The frames having the frame width L set at the interval T are sequentially cut out and supplied to the feature vector generation means 22 at the next stage.
[0010]
The pitch detection means 23 detects the presence or absence of the audio frequency in the frame extracted from the frame extraction means 21, that is, the presence or absence of the pitch, and outputs the detection result to the frame extraction means 21. The frame cutout means 21 changes the shift interval T and the frame width L set according to the detection result by the pitch detection means 23. That is, the shift interval T and the frame width L of the voiced sound section where the voice frequency exists are changed so as to be shorter than the unvoiced sound section where the voice frequency does not exist.
[0011]
The feature vector generation means 22 converts the speech waveform data of each frame that is sequentially input into a feature vector such as a cepstrum coefficient and outputs it to the distance calculation means 24 in the next stage as a feature vector series. . The distance calculation means 24 calculates the distance between the input feature vector series and a speaker model such as a codebook stored in the speaker model storage means 25 provided in the hard disk 5 to recognize the next stage. It outputs to the means 26.
The recognition unit 26 performs speaker recognition by comparing the distance data from the distance calculation unit 24 with a preset threshold value, and displays the result on the display 9, for example.
[0012]
FIG. 3 is a flowchart showing processing by the functional block of FIG. 2. First, in S1, the frame cutout means 21 sets the shift interval T and the frame width L to T ₀ and L ₀ corresponding to the unvoiced sound section. Subsequently, at S2, the frame cutout section 21 a speech waveform data temporarily stored in the buffer memory RAM3 crowded taken from the I / O port 4, the frame width L ₀ which is set in the shift interval T ₀ that has been set Cut sequentially into frames.
[0013]
Analysis processing is performed on the voice waveform data of the cut frame in S3, and pitch detection by the pitch detection means 23 is performed in S4. If pitch detection is not performed, that is, if a speech waveform is not detected, the end of the speech is checked in S5. If not, the processing returns to S1 and the process is repeated.
[0014]
The setting, the pitch detection unit 23 performs pitch detection, in S6, in extracts pitch frequency _{f P,} S7, the frame width L sets the shift interval T to _{T 1} to F _{(f P)} To do. In this case, T ₁ <T ₀ and F (f _P ) <L ₀ . F (f _P) is a function for obtaining the frame width L from pitch frequency f _P, the frame width L when performing pitch detection indicates that changes in accordance with the pitch frequency f _P.
Set the shift interval T to T _1, setting the frame width L to F (f _P), the process is repeated by performing the extraction of the speech waveform returns to S2. When the end of the voice is checked in S5, a recognition process is performed based on the distance for each frame stored in the RAM 3 in advance in S8. After outputting the recognition result, this series of processes is terminated.
[0015]
The analysis process of S3 is the process shown in FIG. That is, a feature vector generation process by the feature vector generation unit 22 is performed in S31, a distance calculation process by the distance calculation unit 24 is performed in S32, and one frame calculated by the distance calculation unit 24 in S33. Each distance is stored in the RAM 3.
[0016]
In such a configuration, when sound is input from the microphone 11, the sound is converted into an electrical analog signal, and the low-pass filter 12 cuts a frequency equal to or higher than a predetermined frequency. In this case, a frequency that is 1/2 or more of the sampling frequency in the A / D converter 13 is cut. For example, if the sampling frequency is 12 kHz, a frequency of 6 kHz or more is cut. The analog signal that has passed through the low-pass filter 12 is sampled at a sampling frequency of 12 kHz in the A / D converter 13 and converted into a digital signal.
[0017]
The voice waveform data converted into the digital signal by the frame cutout means 21 is cut out sequentially into frames having a frame width L at the shift interval T. The presence or absence of a pitch in the frame cut out by the frame cutout means 21 is detected by the pitch detection means 23, and the frame cutout means 21 sets the shift interval T to T ₀ between the state where the pitch is not detected and the state where the pitch is detected. From T to T ₁ , the frame width L is converted from L ₀ to F (f _P ). For example, the shift interval T is converted from 6 msec to 3 msec, and the frame width L is converted from 24 msec to around 12 msec. The reason why the time is around 12 msec is that it varies depending on the pitch frequency.
[0018]
That is, as shown in FIG. 5, with respect to the input speech waveform 100, as shown as frames 101 to 105, a frame is cut out with a frame width of 24 msec while shifting at a shift interval of 6 msec in a section where pitch detection is not performed. When pitch detection is performed, a frame width of 24 msec is converted to about 12 msec as shown as a frame 106, and then a frame with a frame width of about 12 msec is extracted while sequentially shifting at a shift interval of 3 msec.
[0019]
Then, the speech waveform data of the sequentially extracted frames is converted into a feature vector series by the feature vector generation unit 22, and the speaker model stored in the feature vector series and the speaker model storage unit 25 by the distance calculation unit 24. And the distance data calculated by the recognition means 26 is compared with a threshold value. In this way, speaker recognition is performed, and the result is displayed on the display 9.
[0020]
In this way, when the input speech waveform is cut out into frames, the shift interval T and the frame width L are increased in the interval where the pitch is not detected, and after the pitch is detected, the shift interval T and the frame width L are decreased. The number of frames per unit time is relatively greater in the voiced sound interval than in the unvoiced sound interval. That is, the voiced sound section cuts out frames more finely than the unvoiced sound section, and therefore, the specific gravity in speaker recognition becomes large, and speaker recognition with high accuracy can be performed as in the case of weighting.
[0021]
In addition, for the voiced sound section, the shift interval T and the frame width L are reduced to increase the number of frames to be cut out, thereby improving the accuracy. It is not necessary to perform cumbersome weighting, and distance calculation with the speaker model can be easily performed.
[0022]
In addition, since the setting of the frame width L in the voiced sound section is performed using the pitch frequency f _P , that is, the function F (f _P ) that changes according to the frequency of the vocal cord frequency, the frame width L in the voiced sound section is set. It can be set appropriately according to the vocal cord frequency, and the pitch detection accuracy can be maintained regardless of the vocal cord frequency.
[0023]
【The invention's effect】
As described above in detail, according to the present invention, it is possible to provide a speaker recognition device that can perform speaker recognition with high accuracy and can easily calculate the distance from the speaker model.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of an entire apparatus according to an embodiment of the present invention.
FIG. 2 is a block diagram functionally showing the main configuration of the embodiment.
FIG. 3 is a flowchart showing main processing in the embodiment;
4 is a flowchart showing the contents of analysis processing in FIG. 3;
FIG. 5 is a view for explaining frame extraction for an input speech waveform in the same embodiment;
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... CPU, 11 ... Microphone, 13 ... A / D conversion part, 21 ... Frame extraction means, 22 ... Feature vector generation means, 23 ... Pitch detection means, 24 ... Distance calculation means, 25 ... Speaker model storage means, 26 ... recognition means.

Claims (2)

Voice input means for inputting voice, frame cutout means for cutting out voice waveform data input by the voice input means into frames having a frame width set at a set shift interval, and the frame cutout means cut out Feature vector generation means for converting speech waveform data of a frame into a feature vector, pitch detection means for detecting the presence or absence of vocal cord frequency in the frame extracted by the frame extraction means, and speaker model storage means for storing a speaker model in advance And recognition means for performing speaker recognition from the distance between the feature vector generated from the feature vector generation means and the speaker model stored in the speaker model storage means,
When the pitch detection unit detects the vocal cord frequency, the frame cutout unit sets the shift interval to T ₁ (T ₁ <T ₀ ) independent of the pitch frequency, instead of T ₀ corresponding to the unvoiced sound section. A speaker recognition device characterized by that. When the pitch detection means detects the vocal cord frequency, the frame cutout means further sets the frame width to a value smaller than L ₀ that does not depend on the pitch frequency, instead of L ₀ corresponding to the unvoiced sound section. The speaker recognition device according to claim 1.

Images