JPH0646359B2 - Word speech recognizer - Google Patents

Description

Description: TECHNICAL FIELD OF THE INVENTION The present invention relates to a word voice recognition device, and more particularly to a pitch adaptive word such that a voice feature is extracted by a digital filter having a resonance frequency that is a constant multiple of the pitch. The present invention relates to improvement of a method for extracting characteristic parameters in a voice recognition device.

[Prior Art] FIG. 1 is a block diagram showing an electrical configuration of a conventional pitch adaptive word speech recognition apparatus. First, the configuration of a conventional word voice recognition device will be described with reference to FIG. First
In the figure, an audio signal input from a microphone 11 is input to a microphone amplifier 12, amplified, and then given to an AGC circuit 13. The AGC circuit 13 is
The gain of the amplifier provided inside is automatically controlled so that a constant output can be obtained even if the magnitude of the input signal changes. The output of the AGC circuit 13 is given to the A / D conversion circuit 14 and converted into a digital signal.
The output of the A / D conversion circuit 14 is given to the waveform memory 15. The waveform memory 15 temporarily stores one frame of input waveform data. The output of the waveform memory 15 is given to the feature extraction unit 2.

The feature extraction unit 2 includes a pitch period extraction circuit 21, a filter coefficient setting circuit 22, a digital filter 23, a level calculation circuit 25, and a start / end detection circuit 6. The pitch period extraction circuit 21 extracts the pitch frequency in the input speech waveform of one frame temporarily stored in the waveform memory 15. The filter coefficient setting circuit 22 is a pitch period extraction circuit 2
Based on the pitch frequency of the input speech waveform extracted by 1, the filter coefficient is set so that the resonance frequency of the filter is an integral multiple of the pitch frequency. The digital filter 23 is the filter coefficient setting circuit 2
The filter coefficient is determined according to the filter coefficient set in 1. The level calculation circuit 25 calculates the input voice waveform level temporarily stored in the waveform memory 15. The output of the level calculation circuit 25 is given to the recognition processing section 5 and the start / end detection circuit 6. The start / end detection circuit 6 detects the start / end of the input audio signal using the level calculated by the level calculation circuit 25.

The input pattern memory 3 provided in association with the recognition processing unit 5 temporarily stores the characteristic parameters of the voice analyzed by the characteristic extraction unit 2. Further, the registered pattern memory 4 is for storing the characteristic parameter of the registered word or the characteristic parameter of the standard voice analyzed and extracted at the time of registration and giving it to the recognition processing section 5. The recognition processing unit 5 performs recognition processing using the characteristic parameters stored in the input pattern memory 3 and the characteristic parameters registered in advance in the registered pattern memory 4. The recognition processing unit 5 is mainly composed of a microprocessor, for example.

In the word voice recognition device shown in FIG. 1 described above, the voice waveform is divided into frames of a fixed time, and the frequency spectrum at an integer multiple of the pitch frequency is extracted as a characteristic parameter for each frame. During the recognition process, the distance between the frames to be compared is calculated as a numerical value indicating the difference in phoneme between the frame of the registered word and the frame of the input word, and the matching process is performed using this value.

FIG. 2 is a block diagram showing the configuration of the distance calculation unit included in the recognition processing unit 5 shown in FIG. In FIG.
The distance calculation unit 51 calculates a distance between frames in the analyzed voice feature parameter stored in the input pattern memory and the standard voice feature parameter stored in the registered pattern memory 4. The distance calculation unit 51 includes parameter temporary storage memories 511 and 512 and a Chebyshev distance calculation circuit 513. The parameter temporary storage memory 511 temporarily stores one frame of characteristic parameters stored in the input pattern memory, and the parameter temporary storage memory 512 temporarily stores one frame of characteristic parameters stored in the registered pattern memory. It is something to remember. Then, the Chebyshev distance calculation circuit 513 uses the parameter temporary storage memory 511.
And the parameter temporary storage memory 51
The Chebyshev distance with the content temporarily stored in 2 is calculated. The Chebyshev distance calculated by the Chebyshev distance calculation circuit 513 is given to the matching processing unit 52, and the matching processing unit 52 matches the feature parameter of the analyzed voice stored in the input pattern memory 3 with the feature parameter of the standard voice. Do.

Next, the operation of the conventional word speech recognition apparatus will be described with reference to FIGS. 1 and 2. Microphone 1
The audio signal captured at 1 is amplified by the microphone amplifier 12 and given to the AGC circuit 13. The AGC circuit 13 adjusts the maximum value of the input waveform of the audio signal to a constant level, and supplies it to the A / D conversion circuit 14. The A / D conversion circuit 14 encodes the input waveform into a digital signal at each predetermined sampling point. The sampling data for one frame is given to the waveform memory 15 and temporarily stored. The waveform data stored in the waveform memory 15 is input to the level calculation circuit 25 and the pitch period extraction circuit 21. The level calculation circuit 25 calculates the level of the waveform data supplied from the waveform memory 15, and supplies the calculation result to the recognition processing unit 5 and the start / end detection circuit 6.

The start / end detection circuit 6 detects the start and end of the input waveform of the voice signal based on the calculation result from the level calculation circuit 25, determines the voice signal section, and gives the determination result to the recognition processing unit 5. . The pitch cycle extraction circuit 21 extracts the pitch cycle based on the wave data given from the waveform memory 15 to recognize the recognition processing section 5 and the filter coefficient setting circuit 2.
Give to 2 and. The filter coefficient setting circuit 22 sets the filter coefficient in the digital filter 23 based on the pitch cycle calculated by the pitch cycle extraction circuit 21 so that the digital filter 23 has a resonance frequency at an integral multiple of the pitch frequency. The digital filter 23 calculates the frequency spectrum of one frame in the waveform data given from the waveform memory 15 based on the filter coefficient set by the filter coefficient setting circuit 22.

Through the series of operations described above, the feature extraction unit 2 uses the pitch frequency time series [f _pi ], the spectrum time series pattern [Cim], i = 1, 2, ... As the feature parameters for one word.
I), (m = 1, 2 ... M), (I: number of frames of analysis word, M: number of filters for spectrum analysis). The feature parameter for one word thus obtained is stored in the registered pattern memory 4 in the registration mode, stored in the input pattern memory 3 in the recognition mode, and then recognized by the recognition processing unit 5 by the pattern matching method. Perform processing.

As shown in FIG. 2, the recognition processing unit 5 has a distance calculation unit 51 that calculates the distance between the frames of the input pattern and the registered pattern, but the feature parameter stored in the input pattern memory 3 is [f _pi ], [A _im ], and the characteristic parameters of the template stored in the registered pattern memory 4 for performing matching are pitch frequency time series [f _qj ] and spectrum time series pattern [b _jm ].
(J = 1, 2, ..., J), the distance d (i, i between the frame i of the input pattern and the frame j of the registered pattern
j) is Becomes In order to perform this calculation, the matching processing unit 5
2 gives a control signal to the input pattern memory 3 and the registered pattern memory 4, the input pattern memory 3 outputs a _i =
(A _i1 , a _i2 ... a _im ) is the parameter temporary storage memory 51.
1 from the registered pattern memory 4 b _j =
After (b _j1 , b _j2 , ..., B _jm ) is given to the parameter temporary storage memory 512, the Chebyshev distance calculation circuit 51
3 performs the calculation of the above-mentioned equation (1), and d (i, j)
Ask for. The matching processing unit 52 uses d (i, j) calculated by the distance calculation circuit 51 to perform matching processing using a known pattern matching method. Then, the matching distance between the input pattern and the registered pattern is obtained, and the registered pattern having the smallest matching distance is selected as the recognition result.

As described above, in the feature extraction method in the conventional word speech recognition apparatus, the difference between a _im and b _jm is the main value of the distance calculation, but a _im is the frequency spectrum of f _pi × m, and b _jm Is the frequency spectrum of f _qi × m. However, since f _pi and f _qi are not generally equal, spectrums are compared at different frequencies, and if f _pi and f _qj are significantly different, they are unsuitable as feature parameters and the recognition performance is degraded. there were.

SUMMARY OF THE INVENTION Therefore, a main object of the present invention is to set a resonance frequency of a digital filter used for feature extraction to a frequency that is an integral multiple of a pitch frequency that is closest to a predetermined frequency. It is an object of the present invention to provide a word speech recognition device having excellent recognition performance that can effectively extract a spectrum and shorten the recognition processing time.

The above and other objects and features of the present invention will become more apparent from the detailed description given below with reference to the drawings.

[Embodiment of the Invention] FIG. 3 is a schematic block diagram showing an electrical configuration of an embodiment of the present invention. The embodiment shown in FIG. 3 is the same as FIG. 1 described above except that a constant comparison frequency memory 24 is provided in association with the filter coefficient setting circuit 22. The constant comparison frequency memory 24 stores a predetermined frequency for frequency spectrum extraction and supplies it to the filter coefficient setting circuit 22.

FIG. 4 is a diagram showing the result of spectrum pattern extraction by the conventional word voice recognition apparatus shown in FIG. 1 and an embodiment of the present invention.

Next, the operation of the embodiment of the present invention will be described with reference to FIGS. 3 and 4. Since the third embodiment is the same as the conventional word voice recognition apparatus shown in FIGS. 1 and 2 except for the setting of the filter coefficient, detailed description other than the filter coefficient setting operation will be given. Is omitted. When the voice is input and the pitch frequency f _pi which is the output of the pitch period extraction circuit 21 is given to the filter coefficient setting circuit 22, the filter coefficient setting circuit 22 is predetermined from the constant comparison frequency memory 24 for frequency spectrum extraction. Frequency (constant comparison frequency) f _c = f _c1 , f _c2 , ...
f _cN ), (N is the number of filter channels). next, A positive integer k that satisfies
The filter coefficient is set so that the resonance frequency f _{0 of} 3 becomes k · f _pi . This setting is repeated from n = 1 to N, the digital filter 23 receives the filter coefficients,
Spectral pattern a _ci = (a _ci1 , a _ci2 , ... a _ciN )
To extract. The relationship between a _i and a _ci is shown in FIG.

By performing the above calculation for each frame, the feature extraction unit 2 uses a _ci , i = as the feature parameters for one word.
1, 2, ..., I, and f _pi are extracted. At the time of recognition, the distance d (i, j) between the frame i of the input pattern a _ci , i = 1, 2, ..., I and the frame j of the registered pattern b _cj , j = 1, 2 ,. Will be asked for.

[Effects of the Invention] As described above, according to the present invention, since the frequency closest to the fixed frequency (fixed comparison frequency) predetermined by the filter coefficient setting means is set, the pitch frequency changes. However, the resonance frequencies are set relatively close to each other, and the frequency components of the spectrum do not differ greatly for the spectrum of the same number. This is advantageous in the distance calculation between frames in the matching process, and the voice recognition performance is improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing the electrical configuration of a conventional pitch-adaptive word speech recognition device. FIG. 2 is a block diagram showing a configuration of a distance calculation unit in the word voice recognition device shown in FIG. FIG. 3 is a block diagram showing the electrical construction of an embodiment of the present invention. FIG. 4 is a diagram showing a result of spectrum pattern extraction by a conventional word voice recognition device and an embodiment of the present invention. In the figure, 1 is a voice input unit, 2 is a feature extraction unit, 3 is an input pattern memory, 4 is a registered pattern memory, 5 is a recognition processing unit, 6 is a start / end detection circuit, 11 is a microphone, 1
2 is a microphone amplifier, 13 is an AGC circuit, and 14 is A
/ D conversion circuit, 15 waveform memory, 21 pitch period extraction circuit, 22 filter coefficient setting circuit, 23 digital filter, 24 constant comparison frequency memory, 25 level calculation circuit, 51 distance calculation unit, 52 A matching processing unit 511, 512 is a temporary parameter storage memory, 51
3 shows a Chebyshev distance calculation circuit.

Claims (1)

[Claims] 1. A voice signal input means for converting voice into an electric signal for input, a feature extraction means for extracting a feature parameter of a voice signal waveform input from the voice signal input means, and an extraction by the feature extraction means. Input pattern storage means for storing the characteristic parameters of the recognized word voice to be recognized, registered pattern storage means for storing in advance the characteristic parameters of the plurality of word voices extracted by the characteristic extraction means, and the input pattern storage means And a voice recognition processing unit that performs voice recognition processing by calculating the degree of similarity between the feature parameters of the input voice stored in the input pattern storage unit and the feature parameters of the plurality of word voices stored in the registered pattern storage unit. The means is set with a detecting means for detecting the pitch frequency f _p1 of the voice signal input from the voice signal input means. A digital filter whose resonance frequency changes in accordance with the filter coefficient and extracts a plurality (N) of spectrum data of the audio signal as the characteristic parameter; and a plurality of predetermined (N) for extracting a frequency spectrum. ), The fixed comparison frequency storage means for storing the fixed comparison frequencies f _c1 , f _c2 ... F _cN , and the detection of the pitch frequency f _p1 by the detection means, Min | k · f _p1 −f _cn | , (K = 1, 2
...) to calculate the positive integer k _n that satisfies the resonance frequency of the digital filter and a filter coefficient setting means for setting a _{k n · f p1 (n =} 1~N), word recognition device.