JPH0556520B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a speech recognition method, and particularly to a speech recognition method that performs word speech recognition using local peaks.

(Prior art) Conventionally, this type of speech recognition method has been used in the Transactions of the Institute of Electronics and Communication Engineers, J68-A [1] (January 1985) p.78-
There was something listed in 85. Figure 2 is a flowchart of a conventional speech recognition method using local peaks.The input voice is frequency-analyzed every 10 msec by a group of 15-channel bandpass filters (see 1 in Figure 2), and the vocal cord sound source characteristics are As a normalization method for individual differences, the voice spectrum is expressed logarithmically on both the amplitude and frequency axes, and a least squares approximation straight line is obtained (second
(see 2 in the figure), calculate the difference and correct it. However, if the slope of the least squares approximation line is positive, the difference from the average value is taken. After that, as shown in Figure 3, for each frame (10 msec), for each part where the amplitude is 0 dB or more, the channel that has the maximum value among those with an amplitude of 1/2 or more of each maximum value is determined as having a local peak. Binarization is performed by setting one to "1" and the other to "0" (see 3 in Fig. 2). The number of channels of the bandpass filter is 15, but when the slope of the least squares approximation line of the 16th channel is negative, it is regarded as a voiced sound and set to 1, and when the slope is positive, it is regarded as an unvoiced sound and set to "0". Add the slope sign (see 4 in Figure 2).

The weighted average dictionary is obtained as a multivalued pattern by linearly extending and adding a plurality of binarized patterns to the longest one on the time axis (see 5 in FIG. 2).

To match a binary input pattern with a multi-value weighted average dictionary, the longer pattern is linearly stretched and matched in the time direction, calculations are performed based on a certain degree of similarity, and a standard pattern that gives the maximum degree of similarity is selected. The category name is used as the recognition result (see 6 in Figure 2).

(Problems to be Solved by the Invention) However, the conventional methods described above function effectively in environments with a good signal-to-noise ratio, such as when using close-talking microphones, but they do not work particularly well in environments with strong noise. There was a problem in that noise peaks were likely to be picked up as local peaks in silent parts, leading to an increase in erroneous recognition.

The present invention eliminates the problem of conventional technology in which noise peaks are extracted as local peaks of speech, especially in silent parts, in environments with strong noise, resulting in increased erroneous recognition, and provides a speech recognition method that is highly resistant to noise. This is what we provide.

(Means for Solving the Problems) The speech recognition method according to the present invention first performs frequency analysis of input speech into analysis data of a plurality of channels for each speech frame. Further, all analysis data of all voice frames in a predetermined noise section, which is known in advance to be only noise, is averaged to calculate the noise average power. Then, for each audio frame of the input audio, all analysis data corresponding to the frame is averaged to calculate the average power. Next, the average voice power, which is the difference between the above-mentioned average power and the above-mentioned noise average power, is calculated for each voice frame. On the other hand, analysis data of input speech is spectral normalized using a least square approximation straight line of the speech spectrum in the speech frame to which the data belongs. For this spectrum-normalized analysis data, the presence or absence of local peaks along the frequency axis is determined using the aforementioned least squares approximation line as a reference, and if there is a local peak, it is set as "1" and there is no local peak. Conversion to a binarized local peak pattern is performed, with the case of ``0'' being set as ``0''.

Then, the degree of similarity between a plurality of standard patterns represented by local peak patterns prepared in advance and the local peak pattern of the input audio is calculated for each normalized audio frame between the standard patterns and the input audio,
The final similarity is calculated by weighting this similarity using the voice average power described above and accumulating it over all the normalized voice frames described above. As a result, the category name of the standard pattern corresponding to the final similarity that gives the maximum value is determined to be the recognition result.

(Operation) The present invention calculates the average voice power, which is the difference between the average power of the frequency analysis data of the input voice and the noise average power of a predetermined noise section, and compares the similarity between each standard pattern and the local peak pattern of the input voice. Since the final similarity is calculated by weighting the speech average power each time, it is possible to compensate for imperfections in local peak extraction for noise-containing speech, particularly for low-level portions of speech.

(Example) FIG. 1 is a flowchart of the speech recognition method according to the present invention, in which the input speech signal is frequency-analyzed every 10 msec (frame period) in frequency analysis step 11, and multiple channels are transmitted for each speech frame. Logarithmic values are output as analysis data.

Next, from the logarithmic data of the noise section given in advance, in a noise average power calculation step 12, all the logarithmic values of all the speech frames within the noise section are averaged to calculate the noise average power.

On the other hand, in the voice average power calculation step 13, the voice average power is calculated using the difference between the average power found from the frequency-analyzed logarithm value of each voice frame of the input voice and the noise average power.

Further, the logarithmic value of the frequency analysis data of the input speech is spectral normalized using the least square approximation straight line of the speech spectrum in the speech frame to which this logarithmic value belongs (step 14). Next, the presence or absence of a maximum value along the frequency axis, that is, a local peak, is determined for the spectrum-normalized logarithm value using the least squares approximation straight line as a reference, and the presence or absence of a local peak is set as "1", and the local peak is determined. The input local peak pattern is converted into a binary input local peak pattern with "0" in the case of none.

The standard pattern 16 is created in advance by linearly expanding and contracting a plurality of binary local peak patterns to the average utterance length for each category and adding them.

In the matching step 17, the binary input local peak pattern is linearly expanded or contracted to the frame length of each standard pattern, the similarity with each standard pattern is calculated, and the output of the audio average power calculation step 13 is Weight the average power to similarity and accumulate the weighted similarities over all audio frames to obtain the final similarity. Then, the category name of the standard pattern that gives the maximum final similarity is output as the recognition result.

Next, one embodiment of a speech recognition device using the speech recognition method of the present invention will be described.

FIG. 4 is a block diagram showing the overall circuit configuration of one embodiment of the speech recognition device according to the present invention.

In this embodiment, as shown in FIG.
is an audio input terminal, 102 is a frequency analysis section, 103
105 is a spectrum normalization unit, 105 is an average power memory, 106 is a noise power calculation unit, 107 is a voice section detection unit, 108 is a voice average power calculation unit, 109 is a local peak feature extraction unit, 110 is a standard pattern memory, and 111 is a The linear matching section, 112, is a determining section, and 113 is an output terminal for recognition results. The operation will be explained below.

First, an audio signal converted into an electrical signal by a microphone or the like is input to an audio input terminal 101, and a frequency spectrum is extracted by a frequency analysis section 102. The circuit configuration of this frequency analysis section 102 is shown in FIG.

In FIG. 5, 201 is a preamplifier (Amp);
202 is a group of band pass filters (BPF-1,...,
BPF-N), 203 is a full-wave rectifier group (DET-1,
..., DET-N), 204 is a low-pass filter group (LPF-1, ..., LPF-N), 205 is an analog multiplexer, 206 is an AD converter, 20
7 is a ROM for logarithmic conversion. The bandpass filter group has center frequencies arranged at equal logarithmic intervals from 1 to N, starting from the lowest N frequency.
Channel, etc., is written as Nth channel. In this embodiment, N=22.

The output of the low-pass filter group 204 is switched by the multiplexer 205, and the output of the low-pass filter group 204 is switched to the AD converter 2.
In step 06, the signal is converted into a digital quantity at a sampling period of 10 ms, converted into a logarithmic value by the logarithmic conversion ROM 207, and sent to the spectrum normalization unit 103.

In this embodiment, an analog filter is used in the configuration of the frequency analysis section 102, but the structure is essentially the same even if a digital filter bank is used.

The audio spectrum data for N channels of the K-th frame outputted by the logarithmic conversion ROM 207 are assumed to be X ₁ (K), . . . , X _N (K). The spectrum normalization unit 103 first calculates the average value (K) for each frame based on the following equation (1).

(K)= _N 〓 ⁱ⁼¹ X _i (K)/N...(1) Hereinafter, (K) will be referred to as "average power" in K frames. Average power (K) is average power memory 105
At the same time, it is sent to the noise power calculation section 106.

The noise power calculation unit 106 calculates the average noise power in a section [K _E1 , K _E2 ] that is known to be only noise in advance based on an instruction input from the outside.
The instructions for the noise-only section shall be known in advance by asking the speaker to stop speaking.

For noise power measurement, the noise interval [K _E1 , K _E2 ]
The noise average power N _p at is calculated as shown in the following equation (2).

N _p = 1/K _E2 −K _E1 +1 _KE2 〓 ^K=KE1 (K) …(2) In this example, K _E2 −K _E1 = 20 to 30 (frames)
It is said that As the next step, the voice section detecting section 107 detects the voice section.

Speech section detection is performed using, for example, two threshold values E ₁ and E ₂ (E ₁ >E ₂ ) for the average power. That is, as shown in FIG. 6, when the average power exceeds E ₂ and then exceeds E ₁ without becoming smaller than E ₂ , the point where E ₂ is exceeded is defined as the starting point K ₁ . In the same way, the time axis is reversed to find the end point K ₂ , and [K ₁ , K ₂ ] is defined as the speech interval.

The thresholds E ₁ and E ₂ are determined from the noise average power N _p by E ₁ = N _p + L ₁ E ₂ = N _p + L ₂ (3) where L ₁ and L ₂ are arbitrary constants, and L ₁ > L ₂ > 0. It is. It is said that

In addition, the spectrum normalization unit 103 generates audio spectrum data X _i (K) (K=K ₁ ,...,K ₂ ; i=
1,...,N) is spectral normalized using the least squares approximation line according to the following equation (4).

Y _i (K)=A(K)・i+B(K)...(4) In other words, calculate the difference between the audio spectrum data X _i (K) and the least squares approximation straight line using the following equation (5),
Obtain spectral normalized data Z _i (K).

Z _i (K) = X _i (K) - (A (K) · i + B (K)) (K = K ₁ , ..., K ₂ ; i = 1, ..., N) ... (5) Above (4) A(K) on the least squares approximation straight line of Eq.
B(K) is given by the following equations (6) and (7).

The above equation (5) is calculated when A(K) is positive, that is, in most cases, even for unvoiced sounds. In the time direction, calculations using equation (5) are performed for all frames K before the voice section is detected. Also, the average voice power (K) is calculated from the average power (K) of the voice section, K _E [K ₁ , K ₂ ], and the noise average power N _p .

′(K)=(K)−N _p (however, (
K)>N _p ) 0 (However, if (K)≦N _p ) (K=K ₁ ,...,K ₂ )...(8) The above equation (8) is calculated by the audio average power calculation unit 108 It will be held in In other words, ′(K) is max{0,
X(K)−N _p }.

Next, the local peak feature extraction unit 109 extracts the spectrum normalized data from the start to the end of the voice.
Z _i (K), (K=K ₁ , . . . , K ₂ ; i=1, . . . , N) are sequentially processed for each frame.

The detailed configuration of the local peak feature extraction section 109 is shown in FIG.

In FIG. 7, 401 is a data storage memory;
402 is a peak extraction unit, and 403 is a local peak pattern storage memory.

The data storage memory 401 has a starting point _K1 to an ending point.
Spectral normalized data Z _i (K) up to K ₂ (K=
K ₁ ,...,K ₂ ; i=1,2,...,N) are stored, and spectrum normalized data Z _i (K) (K=K ₁ ,
..., K ₂ ; i=1, ..., N) local peaks are detected. This peak extraction is performed using spectral normalized data Z _i (K) (K=K ₁ ,..., K ₂ ; i
=1,...,N) as follows.

Z _i+1 (K)−Z _i (K)＞0 …(9) Z _i+2 (K)−Z _i+1 (K)＜0 …(10) Expressions (9) and (10) above When both are satisfied at the same time, the spectral normalized data Z _i+1 (K) is the maximum value, and (i
+1) indicates a channel number with a local peak, and the local peak pattern P _i+1 (K) is set to 1.

The spectrum normalized data from i=1 to (N-1) is sequentially checked to see if it satisfies equations (9) and (10), and if it is, the local peak pattern P _i+1 is determined.
(K) is set to "1", and if it is not satisfied, it is set to "0" and binarized and stored in the peak pattern storage memory 403.

However, as a boundary condition, when i = 1 and Z _{i +1} (K) - Z _i (K) < 0, P ₁ (K)
=1...(11) When i=(N-1) and Z _i+1 (K)-Z _i (K)>0
P _N (K)=1 (12), and if the condition is not met, it is set to "0", binarized, and stored in the local peak pattern storage memory 403.

The local peak pattern in a certain K-th frame in the local peak feature extraction unit 109
P _i (K) (i=1,2,...,N; K= _K1 ,..., _K2 )
A flowchart of the extraction is shown in FIG.

In this way, the spectral normalized data Z _i (K) (i=1, 2,...,
N; K=K ₁ , . . . , K ₂ ) are read out for each frame to obtain a binary local peak pattern.

The standard pattern memory 110 stores a standard pattern created for each category by overlapping a plurality of local peak patterns determined in advance. In order to normalize the difference in speech rate, the standard pattern calculates the average utterance length for each category, and linearly expands and contracts the time-direction frame length of the binary local peak pattern to match the average utterance length, from 1 to N channels. It is created as an overlapping, multivalued pattern. The standard pattern memory 110 also stores a standard pattern length SL corresponding to each standard pattern.

The linear matching unit 111 linearly expands and contracts the time frame length of the input local peak pattern input from the local peak feature extraction unit 109 to the standard pattern length SL of each category in the standard pattern memory 110, and then Initial similarity calculation and weighting of similarity based on average power are performed for each time.

The initial similarity S^(K) is calculated for each frame based on the following equation (13).

S^(K)=P D ^t /‖P‖ ‖D‖ (13) Here, P indicates the feature vector of the input local peak pattern in the K-th frame.

P=(P ₁ (K), P ₂ (K), ..., P _N (K)) ... (14) D indicates a feature vector in the K-th frame of an arbitrary standard pattern.

D=(D ₁ (K), D ₂ (K), ..., D _N (K)) ...(15) D ^t represents transposition, and ‖P‖ and ‖D‖ are respectively P,
represents the norm of D.

Furthermore, when the average power of the voice is small, it is easily affected by noise, so the final similarity S is determined from the following equation (16).

Here, ′(K) is the average voice power found by equation (8) above, ^max _K {′(K)} is the maximum value of the average voice power in the time-normalized interval of the input audio, and SL is the standard The pattern is long. In other words, the final similarity S is the value obtained by weighting the initial similarity S^(K) obtained by the above equation (13) with the speech average power X'(K) obtained from the above equation (8), which is obtained by subtracting the noise average power N _p It is.

The configuration of the linear matching section 111 is as shown in FIG. 9, where 701 is an input terminal for an input local peak pattern, 702 is an input terminal for a standard pattern, and 703 is an input terminal for an input local peak pattern.
is an input terminal for audio average power, 705 is an input terminal for standard pattern length, 706, 707, 708 are multiplication adders, 709, 710 are square root operators, 711 is a multiplier, 712 is a divider, 713 is a multiplication adder ,
714 is a multiplier, 715 is a maximum value detector, 716
is a multiplier, and 717 is an output terminal.

The norm of the input local peak pattern P input from the input terminal 701 is calculated by the multiplier adder 706 and the parallel root operator 709 as shown in the following equation (17). Similarly, the norm of the standard pattern D is determined by the multiplier/adder 708 and the square root calculator 701. Further, P·D is determined by the multiplier/adder 707 as shown in the following equation (18). _N 〓 ⁱ⁼¹ P _i (K)·D _i (K) (18) Furthermore, the divider 712 determines the initial similarity S^(K) of the above equation (13).

On the other hand, the average voice power '(K) is at the input terminal 70.
3, and is determined by the multiplier/adder 713 as _SL 〓 ^K=1 S^(K)·'(K) (19). In addition, the maximum value detector 715 determines the maximum value max{'(K)} K _E [1, SL] ...(20) of the time-normalized audio average power, and the multiplier 716 determines the standard pattern length.
This result and the result of equation (19) are input to the divider 714, and the final similarity S is determined.
It is output to the output terminal 717 and sent to the determination section 112.

The determining unit 112 outputs the category name or category number of the standard pattern that gives the maximum final similarity value for each standard pattern via the output terminal 113, as a recognition result.

(Effects of the Invention) As described in detail above, in the present invention, as a local peak extraction method, a least squares approximation straight line is used as a reference regardless of whether it is a voiced sound or an unvoiced sound, or a voice average power (=average power – noise average power)
Since the weighting is performed by weighting, it is possible to compensate for the incompleteness of the local peak extraction of the voice, especially in the low-level part of the voice, even if the voice is mixed with noise.In other words, the peak due to the noise is considered to be the local peak due to the voice. There are few problems and the effect is well recognized.

[Brief explanation of drawings]

Fig. 1 is a flowchart of a speech recognition method according to the present invention, Fig. 2 is a flowchart of a conventional speech recognition method, Fig. 3 is a diagram for explaining binarization of an input speech signal, and Fig. 4 is a diagram of the present invention. A block diagram showing the overall circuit configuration of one embodiment of the speech recognition device according to the invention,
FIG. 5 is a block diagram showing the circuit configuration of the frequency analysis section, and FIG. 6 is a diagram for explaining voice section detection.
FIG. 7 is a block diagram showing the configuration of the local peak feature extraction section, FIG. 8 is a flowchart of local peak pattern extraction, and FIG. 9 is a block diagram showing the circuit configuration of the linear matching section. 101...Input terminal, 102...Frequency analysis section, 103...Spectrum normalization section, 105...
Average power memory, 106...Noise power calculation unit, 1
07...Speech section detection section, 108...Speech average power calculation section, 109...Local peak feature extraction section, 110...Standard pattern memory, 111...Linear matching section, 112...Judgment section, 113...
Output terminal.

Claims (1)

[Scope of Claims] 1. A process of frequency-analyzing input audio and extracting analysis data of multiple channels for each audio frame, and averaging all the analysis data of all audio frames in a predetermined noise interval to obtain a noise average power. A process of calculating the average power by averaging all the analysis data corresponding to the frame for each audio frame of the input audio, and calculating the average power and the noise average for each audio frame of the input audio. A process of calculating the voice average power which is the difference from the power; a process of spectral normalizing the analysis data of the input voice using a least squares approximation straight line of the voice spectrum in the voice frame to which the data belongs; The presence or absence of a maximum value along the frequency axis, that is, a local peak, is determined for the analyzed data using the least squares approximation line as a reference, and the presence of a local peak is set as "1", and the case of no local peak is set as "0". ”, converting the binarized input audio into a local peak pattern, and converting the similarity between the local peak pattern of the input audio and a plurality of standard patterns represented by local peak patterns prepared in advance to the standard pattern. and the final similarity is calculated for each normalized audio frame between the input audio, weighting the similarity by the audio average power, and accumulating over all the normalized audio frames. and a process of determining, as a recognition result, a category name of a standard pattern corresponding to the maximum value of the final similarity for each standard pattern.