JPS61275899A - Voice recognition equipment - 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 [Field of Industrial Application] The present invention relates to a speech recognition device using a bandpass filter bank in an acoustic analysis section.

[Summary of the invention]

This invention uses a pan 1-pass filter bank in the acoustic analysis section of a speech recognition device, and features features of both the mel scale, which corresponds to auditory characteristics, and the commonly used log scale. It is a Hunt bass filter bank with +1 h2, with equal intervals on the low frequency side (51 mel scale, and 5 on the high frequency side).
It is constructed by dividing the frequency at equal intervals on the 1G scale. As a result, there is an advantage that the number of channels in the filter bank can be reduced and the filter installation d1 can be facilitated.

[Conventional technology]

Sound is a phenomenon that changes along the time axis, and Spectra J8
・By uttering sounds in such a way that the patterns change from moment to moment, 41. be caught. Speech recognition is a technology that automatically recognizes short stories and words uttered by humans, but it is currently extremely difficult to achieve speech recognition that is comparable to the human auditory function. For this reason, most of the audio recognition systems currently in practical use perform pattern matching between the standard pattern of the eulogy to be recognized and the input pattern under a certain usage condition ('+). This method is done by

FIG. 1 is a block diagram of an example of this speech recognition device, in which speech input from a microphone f11 is supplied to an acoustic analysis circuit (2). This acoustic analysis circuit (2) extracts acoustic parameters representing the characteristics of the input speech pattern. Various acoustic analysis methods can be considered to extract these acoustic parameters, but one example is to use a bandpass filter and a rectifier circuit as one channel, and to divide such a channel into multiple channels each having a passband that is obtained by dividing the audio band. A method is used in which a temporal change in a spectrum pattern is extracted as the output of this group of band-pass filters.

That is, in the acoustic analysis circuit (2), an audio signal from the microphone (1) is supplied to an A/D converter (213) via an amplifier (21+) and a band-limiting low-pass filter (212). .5kll
It is converted into a 12-bit digital audio signal at a sampling frequency of z. This digital audio signal consists of one row of
Digital bandpass filters for each channel of a 6-channel bandpass filter bank (221+)
, (2212) , ..., (221
+ et al.). This digital band pass filter (22h), (2212),...
(22L+g) is composed of, for example, a Butterworth 4th order digital filter, and is 25011z to 5.5KIIz.
Each band obtained by dividing the band up to 100 nm at equal intervals on the logarithmic axis becomes the pass band of each filter. That is, a bandpass filter bank for 16 channels is constructed by dividing the frequency at equal intervals on a 1G scale.

Each digital bandpass filter (22h L.

The output signals of (2212), ..., (22L+6) are rectifier circuits (222) and (222
2) ,...

(22216) and these rectifier circuits (222+
).

(2222>, ...(2221G) outputs are each digital low-pass filter (223+)
, (2232).

..., (2231e). These digital low-pass filters (223+), (22
32) ,...

(22316) is, for example, a cutoff frequency of 52.811
It is composed of a Z FIR low pass filter.

Each digital low-pass filter (223t), (2232), ... which is the output of the acoustic analysis circuit (2)
.

The output signal of (223z6) is supplied to a sampler (23) constituting the feature extraction circuit (23). This sampler (231) uses a digital low-pass filter (22
31) , (2232) ,..., (22
31G) is sampled every frame period of 5.12m5ec. Therefore, from this, a sample time series At(nl (1=1.2.-16; n is the frame number and n=1.2. . . , N) is obtained.

The output from this sampler (231), that is, the sample time series At1n+, is supplied to a sound source information normalization circuit (232), which removes differences in vocal cord sound source characteristics depending on the speaker of the speech to be recognized.

That is, 81 (nl=
log(^1(nl+B)
...-(tl logarithmic transformation is performed. In this equation (11), B is set to a value such that the noise level is hidden by the bias.

Then, when the vocal cord sound source characteristics are approximated by the formula yi=a−i+b, the coefficients of a and b are determined by the following formula.

(1 = 16> ...(2) (I = 16
)...(3) Then, if the normalized parameter of the sound source is Pifnl, a (n) < Q
Then the parameter Pi(n) is Pi(nl-At(nl
−(a fnl ・i → −bfnll ・−−
It is expressed as T41.

Also, a (when nl≧0, only level normalization is performed,
The parameter Pifnl is expressed as...(5).

In this way, the acoustic parameter time series Pifnl in which differences in vocal cord sound source characteristics are normalized and removed is obtained from this sound source information normalization circuit (232>).

The acoustic parameters Pi (nl) from this sound source information normalization circuit (232) are supplied to the voice interval parameter memory (8). In response to the speech section determination signal, a parameter Pifn) is stored for each determined speech section.

The voice section determination circuit (24) includes a zero cross counter (24).
1), a power calculation circuit (242), and a voice section determination circuit (
243), and the digital audio signal from the A/D converter (2]3) is supplied to the zero-cross counter (241) and the power calculation circuit (242).The zero-cross counter (241) has an I frame period of 5. 12m5e
c, the number of zero crosses of the 64 samples of the digital audio signal within this one frame period is counted, and the count value is supplied to the first input terminal of the audio section determination circuit (243).Power calculation circuit (242), the power of the digital audio signal within this l frame period, that is, the sum of squares, is determined for each frame period, and the output power signal is sent to the second voice section determining circuit (243). The voice section determining circuit (243) is further supplied with voice source normalization information from the voice source information normalization circuit (232) at its third input terminal. In the section determining circuit (243), the number of zero crossings, the power within the section, and the sound source normalization information are processed in a composite manner, and a process of determining silence, unvoiced sound, and voiced sound is performed, and a voice section is determined.

The voice interval determination signal indicating the determined voice interval from the voice interval determination circuit (243) is sent to the voice interval determination surface 178.
The output of (24) is the voice interval parameter memory (
200).

In this way, the memory (200) is
The acoustic parameter time series Pifnl stored in is NA
T treatment episode V! 1f91 will be paid in advance.

The NAT processing circuit (9) includes a trajectory length calculation circuit (91), an interpolation interval calculation circuit (92), and an interpolation point extraction circuit (93).

Parameter time series Pifn) from parameter memory (200) (i = 1. 2. −, 16; n
=1. 2゜..., N) is the trajectory length calculation circuit (91)
supplied to

In this trajectory length calculation circuit (91), the acoustic parameter time series Pi (nl) calculates the length of the trajectory by linear approximation drawn in the parameter space as shown in FIG. This is the position taken by the value of each parameter (this is shown as a case of a two-dimensional space of Pl and P2 for the sake of explanation).

In this case, the Euclidean distance D (at + bI) between the ■-dimensional vectors aI and bI is ■D(at, bI)-Σ (a(-bI)2i +1 ... (6). Therefore, the ■-dimensional The acoustic parameter time series Pi(
From nl, the distance S between adjacent parameters in the time series direction when the trajectory is estimated by linear approximation (nl is 5(nl
=D (PI(n+1), PiTn))(n=
1. ..., N) ...(7). Then, the distance from the first parameter Pi(1) to the n-th parameter Pifnl in the time series direction is expressed as 1111 S fun). In addition, St.
, (11-0.

The total track length m SL is the track length SL calculated by this track length calculation circuit (91).
A signal indicating the interpolation interval calculation circuit (92) is supplied to the interpolation interval calculation circuit (92).

This interpolation interval calculation circuit (92) calculates the resampling interval T when resampling is performed along the locus.

In this case, if resampling is performed at M points, the resampling interval T is T=SL/ (M-1) −・=
It is determined as l+01.

A signal indicating the 100 sampling interval T from this interpolation interval calculation circuit (92) is supplied to an interpolation point extraction circuit (93). In addition, the acoustic parameter time series Pi (nl from the parameter memo January 8) is also obtained from this interpolation point extraction circuit (9
3). This interpolation point extraction circuit (93) resamples the acoustic parameter time series Pifnl at a resampling interval T as shown by the circle in FIG. Then, the recognition parameter time series old hl is formed from the new point sequence obtained by this zumbling.

Here, the interpolation point extraction circuit (93) performs processing according to the flowchart shown in FIG. 2 to form the recognition parameter time series old fml.

First, in step (101), the value 1 is set to the variable 1 indicating the number of the resampling point in the time series direction, and the value 1 is set to the variable IC indicating the frame number of the acoustic parameter time series Pi (nl). , is initialized.Next, in step (102), the variable ■ is incremented, and in step (1031), the variable at that time, 1
It is determined whether or not is (M-1) or later, and the number 4J in the chronological direction of the sampling point at that time is the last number that needs to be renumbered. Check to see if it is. If the number is the highest, the process advances to step (104) and the nail sampling ends.

If it is not the last number, step [05] is the first resampling point (this is always a silent part).
The resampling distance DC from to the first resampling point is calculated. Next, the process proceeds to step (106), where the variable IC is incremented. Then step (
107), the distance s+,+ from the first parameter gate (1) of the acoustic parameter time series Pifnl to the 1Cth parameter Ptac+ is
. Depending on whether the resampling point is smaller than the parameter P i on the trajectory, the resampling point at that time
It is determined whether the position is closer to the start point of the trajectory than +1o, and if it is not located to the start point side, the process returns to step (106) and the variable IC is incremented (step (moxibustion) again).
107), 'C resampling point and parameter P i
ac), and when it is determined that the resampling point is located closer to the starting point than the parameters Pi, 1o on the trajectory, step (108)
Then, the recognition parameter (old σ) is formed.

That is, from the sampling distance obtained by the fifth resampling point, the (1F, -1)th parameter P located closer to the starting point than this fifth resampling point
Subtract the distance 5LfIC-11 due to 1ac-n to get the (
The distance MSS from the IC-1)th parameter P L (Ic-11 to the first resampling point is determined).

Next, the distance 5fnl (this distance S t
n+ is obtained by signal processing shown in equation (7). )
Divide this distance MSS by and get the division result SS/S
(In Ic-11, parameters Pioc+ and P i located on both sides of the fifth resampling point on the trajectory,
The difference from 1 cm11 (P i uc, P i (ic-1
, ) in 11), and calculate the parameter P of the (Ic-1)th L1 located adjacent to the starting point side of the fifth resampling point on the trajectory -.
Calculate the interpolation effect from L+c-11, and use this interpolation M and the fifth
The (IC-1)th parameter P i uc-1+ located adjacent to the starting point side of the nail sampling point
and a new recognition parameter along the trajectory old (1
)) is formed.

In this way, U7 excludes the start point and end point (when they are silent, they are old (Il = PHol " 0, old fMl - gate IG) - 0) < (M-2)
The recognition parameter time series old (m
l is formed.

The recognition parameters from this NAT processing circuit (9) are stored in the standard pattern memo (January 4) for each journal target word in the registration mode in the chronological order (ml is the mode 1 changeover switch (3)). be done. Further, in the recognition mode, the distance fIVJJ is supplied to the distance fIVJJ calculation circuit (6), and the distance between the parameter time series of the standard pattern from the standard pattern memory (4) is calculated. The distance in this case is calculated, for example, as a simple Chech u'j-fu distance. Each I! from this distance calculation circuit (0)! ! The calculated output of the distance between the quasi pattern and the input cover pattern is supplied to the minimum value determination circuit (7), and the standard pattern with the minimum calculated distance value is determined.
Based on this judgment result, the recognition result of the input voice is transferred to the output terminal (70
) can be obtained.

Note that in the above example, the time normalization process and Step 17 estimate the trajectory that the acoustic parameter time series Pifnl draws in its parameter space, and resample along that trajectory to obtain a new normalized parameter time series old+m+. However, other methods may be used for time iE normalization, such as a method that performs so-called DP matching when calculating distance.

[Problem that the invention seeks to solve]

As described above, a bandpass filter bank is often used in the acoustic analysis section, and in that case, as mentioned above, conventionally, the frequency axis is set on a log scale, and the audio bands are equally spaced on the frequency axis. It is divided into multiple channels.

By the way, when the audio band frequency is divided into equal intervals on a log scale like this, the low range becomes fine and the high range becomes coarse. For this reason, the frequency resolution of the acoustic analysis unit in speech recognition has a large number of low-frequency channels and a small number of high-frequency channels, making it difficult to construct an optimal filter bank with a small number of channels.

Another problem is that the number of channels in the low range increases, making it difficult to design each filter.

[Means for solving problems]

In this invention, the bandpass filter bank of the acoustic analysis section is constructed by dividing the audio frequency band into equal intervals on the mel scale on the low frequency side and equal intervals on the log scale on the high frequency side.

[Effect]

By dividing the low-pass band at equal intervals on the mel scale, the number of channels on the low-pass side is smaller than when dividing the band at equal intervals on the log scale, and it is possible to reduce the number of channels for the filter bank as a whole. It also makes filter design easier.

[Example] 1 when configuring a 16-channel bandpass bank
Indicates a column.

In this example, the low frequency side is divided into 5 channels and the high frequency side is divided into 11 channels.

The frequency band to be divided into five channels on the low frequency side is 0.25 kHz to 0.85 kllz, and this is divided at equal intervals on the mel scale.

Further, the frequency band divided into 11 channels on the high frequency side is set to 0, 85kllz to 5, 2kllz, and this is divided at equal intervals on a log scale.

In this case, a fourth-order Butterworth bandpass filter a1 is installed so that the passbands of adjacent channels cross each other at a point of -3 dB in the entire passband.

The approximate formula for the value X on the mel scale with respect to the frequency f is x = log2 (f/1000+1)
It is expressed as f ≦850 (11), and the conversion formula between the frequency f and the value y of log scale - L is V = log2f (f > 850)
---(12).

The passing center frequencies of each channel of the 16-channel bandpass filter designed as described above are shown in the following table.

〔Effect of the invention〕

The Mel scale corresponds to the characteristics of human hearing, and is coarser in the low range and finer in the high range compared to the log scale.

According to this invention, by configuring the low frequency side using Mel scale and the high frequency side using log scale, the number of filters on the low frequency side can be reduced, and as a result, the number of channels in the filter bank can be reduced. This also makes it easier to design the filter as a whole since the number of low-pass filters is reduced.

[Brief explanation of drawings]

FIG. 1 is a block diagram of an example of a speech recognition device, and FIGS. 2 to 4 are diagrams for explaining the same. (2) is the acoustic analysis section, (221z) to (22116
) are each filter in the bandpass filter bank. :, 1. Il 8 to 1 ~---------" 〇- 6〕!

Claims (1)

[Claims] Input speech is supplied to an acoustic analysis unit having a band-pass filter bank to obtain acoustic parameters, and the difference between the input acoustic parameters and the acoustic parameters of registered standard patterns of recognition target words is calculated. In a device that performs speech recognition based on the calculated output, the above-mentioned band-pass filter bank is configured by dividing the frequency into equal intervals on the mel scale for the low frequency side and equal intervals on the log scale for the high frequency side. A voice recognition device.