JPS61206000A - 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 Application of the Invention] The present invention relates to a speech recognition device, and in particular, to a speech recognition device that extracts a stable spectrum pattern that is not affected by fluctuations in speech level, and that alleviates the effects of fluctuations in the fundamental frequency of speech and analysis. The present invention relates to a speech recognition device.

[Background of the invention]

The conventional spectral pattern normalization method consisted of two-stage processing (level correction, reference level matching) as shown in Figure 1 (Reference ``Simplified Speaker-Independent Speech Recognition'',
Speech Research Group Material No. 884-152 Acoustical Society of Japan, etc.)
. However, although the conventional method has the advantage of performing normalization that maintains the absolute level of the original spectral pattern to some extent, it does not take into consideration the effective use of the calculation length of analysis due to the difference in absolute level caused by the vocalization level. It wasn't. With a speech recognition device that inputs discrete utterances that can adjust the utterance level to some extent (strictly speaking, a speech recognition bag fi!E that inputs input using the same utterance method as the voice to be registered), the conventional method does not pose much of a problem. Ta. However, in continuous voice recognition devices where the voice to be registered and the voice to be input are uttered in different ways, it is difficult to determine where the voice to be registered (generally a part of the input voice) appears in the continuously uttered voice.
For example, the absolute level of the voice may vary greatly depending on whether it is in the middle or at the end of a word), and when the absolute level is small, the computational length of the analysis cannot be used effectively. Furthermore, in conventional band-pass filter (BPF) analysis, the shape of the spectral pattern differs from the original one due to the influence of harmonics of the fundamental frequency of the voice and the influence of analysis fluctuations due to insufficient calculation accuracy of the analysis filter (for example, In some cases, patterns (such as local peaks appearing in the low frequency band) were extracted.

[Purpose of the invention]

An object of the present invention is to provide a pattern normalization method for a speech recognition device in which the calculation length of analysis is effectively utilized even if the absolute level of speech differs, and which is unaffected by the influence of the fundamental frequency of speech or the fluctuation of analysis. It's about doing.

[Summary of the invention]

In order to achieve the above object, the first feature of the present invention is to perform level normalization in which the spectral value of each band is divided by a value corresponding to the relaxation of the value in each band of the short-time spectral pattern. The second feature is that the spectrum pattern is smoothed.

[Embodiments of the invention]

The principle of the present invention will be explained in detail below.

FIG. 1 shows the flow of conventional pattern normalization processing. In the first step, the level is corrected by logarithmic transformation, etc., using as input the spectral pattern fi,i''0 to N channels obtained as a result of band-pass filter (BPF) analysis.This is based on the human auditory characteristics. Specific examples of logarithmic transformation include those shown in the flowchart, as well as the simple logarithmic transformation of a fixed-point value object described in the above-mentioned document "Simplified speaker-independent speech recognition." etc. can also be considered. Next, in the second step, reference level matching is performed to normalize the difference in absolute level caused by the difference in utterance level.

A flow example of the pattern normalization process according to the present invention is shown in FIG. 2, and each step will be explained in detail below.

Input; spectral pattern f,, i = 1 to N channels (1) 1st step (level normalization) f,' = f,
/f, , , where f, , -=Σf.

Level normalization is performed for the purpose of making effective use of the calculation length for one analysis without being affected by differences in absolute levels. As a result, the logarithmic conversion process in the second step can be executed with high accuracy even for voices having a small absolute level.

(2) Second step (level correction) fl' == nogxa (1,0+f1/f,
), where f, ; constant This process is equivalent to the first step of the conventional method.

(3) Third step (standard level adjustment) f, t″=
= f, # f, , , #.

However, f,,'=(Σf,')/N is equivalent to the second step of the conventional method.

(4) Fourth step (smoothing process) f,'=f1"'-1+2ft"+f1.1"')/4
However, i = l~n (N) In general, the center frequency arrangement for bandpass filter analysis is based on the human auditory characteristics and the distribution of the first formant frequency of vowels (for Japanese 5 vowels (male), Ial is 600~900 H
z, Ial is 160-300Hz, jul is 200-5
00H2, lel is 300-600 Hz, 101 is 4
00 to 650 Hz), it is customary to use a logarithmic scale or mel scale arrangement.

In such cases, the center frequency tends to be placed extremely low frequency side (for example, in the band 150-4000H).
When 16 channels are arranged for z and the center frequency is arranged on a logarithmic scale, 6 channels are arranged for frequencies below 500 Hz, and 10 channels are arranged for frequencies below IKHz). As a result, the influence of harmonics of the fundamental frequency of speech (the periodic part of the speech waveform, corresponding to the vocal fold frequency, 100 to 150 Hz for men) that is not related to phonological characteristics (e.g. fat, fil, etc.) is reduced. Smoothing processing is performed to deal with this phenomenon, which tends to appear in the next channel and may cause the spectrum pattern to collapse (for example, a few small peaks appear in the low order), especially around 500Hz. Perform for low-order channels up to The processing may be a moving average of three channels as in the example of the present invention.

FIG. 3 is a diagram showing a comparison of spectral patterns obtained by the conventional pattern normalization method and method 1 of the present invention (from the first step to the third step). The audio is l aikurus
Words with significantly different absolute levels from hi I 1i+
The result of method 1 of the present invention, which inputs the word ending fil, is that the peak (first formant) is sharp without being crushed, and the absolute level of the word ending fil is small, compared to the conventional method.
The dynamic range ratio of the pattern is .

The conventional method (W,) has 2.4 countries, which is an increase of 1.4 times, and is a good pattern normalization.

Figure 4 shows the effect of method 2 of the present invention (processing from the first step to the fourth step including smoothing processing). It can be seen that the small peaks have disappeared. The top pattern is the LPG spectrum and represents the original spectral pattern.

Next, specific embodiments of the present invention will be described in detail.

FIG. 5 is a block diagram showing the configuration of an embodiment of a speech recognition device using the present invention. Input audio 1 is a low-frequency r-wave generator (L
PF), and an analog-to-digital converter (ADC) 2 samples the analog value into a digital value while removing aliasing noise. Thereafter, the input voice is analyzed by the voice analysis section 3 to obtain a spectrum pattern necessary for recognition.

Possible methods for determining the spectral pattern include, for example, band pass filter (BPF) analysis and FET analysis. Thereafter, the pattern according to the present invention is stored in the standard speech storage section 5 as the zero-order normalized input spectrum pattern, in which the normalization section 4 normalizes and corrects the voice utterance level fluctuations and bandpass filter analysis results. The distance calculator 6 calculates the difference (or degree of similarity) between the spectrum pattern and the standard speech spectrum pattern. As the distance, relaxation of the absolute value of the difference in spectral values between channels can be considered.

Next, the matching section 7 calculates the total distance (matching) value including the difference in time structure between the input speech and the standard speech, and the determining section 8 calculates the input speech based on the magnitude relationship of the total distance value. A determination is made as to which standard voice is most similar, and a recognition result 9 is output. The distance calculating section 6 is composed of only simple adders and subtracters, and the matching section 7 is, for example, a continuous NL (Non tan 1 raar).
Matching method (known example: continuous DP method, JP-A-55-22
The 6-judgment section 8, which is constructed from a circuit according to the improvement of the No. 05 publication), can also be constructed from a simple large Ij% comparator. Further, the division and logarithmic conversion processing in the pattern normalization unit 4 of the present invention are executed by table look-up processing, and the rest may be configured only by adders and subtracters.

〔Effect of the invention〕

FIG. 6 shows the results of consonant recognition of continuous speech by the conventional method and the method of the present invention, and shows the recognition rate when multiple candidates up to the nth rank are allowed. As a result, in the case of method 2 of the present invention, which includes smoothing processing, compared to the conventional method, an improvement in recognition rate of about 3% (within the first place) was achieved. The improvement is particularly remarkable when looking at the top 4 and 5 rankings.

As described above, according to the present invention, the calculation length of the analysis can be effectively used regardless of absolute level fluctuations, and the fundamental frequency in the band-pass filter analysis can be used to reduce the influence of fluctuations in the analysis. This has the effect of improving performance.

[Brief explanation of drawings]

Fig. 1 is a diagram showing the flow of Rinrai's pattern normalization processing, Fig. 2 is a diagram showing an example of the flow of pattern normalization processing of the present invention, and Fig. 3 is a diagram showing the conventional method and the method of the present invention. FIG. 4 is a diagram showing the effect of the smoothing processing method of the present invention, FIG. 5 is a block diagram showing an embodiment of a speech recognition device incorporating the present invention, and FIG. The figure is a diagram showing the results of a recognition experiment showing the effects of the present invention. 4...Pattern normalization section. Y j 口 2nd 'tJ J 口(a) 硅明一个A; 1 (b) Subjective/method 11"tz<之' 5.

Claims (1)

[Scope of Claims] 1. In a speech recognition device equipped with a speech analysis section for extracting at least a short-time spectral pattern of speech, a first means for normalizing the level of the spectral pattern obtained by the speech analysis section; , a second means for level correction conversion of the pattern obtained by the first means, and a third means for adjusting the reference level of the pattern obtained by the second means. speech recognition device. 2. The speech recognition device according to item 1, further comprising means for smoothing the spectrum pattern obtained by the speech analysis section using spectrum information of adjacent bands. 3. In the speech recognition device according to item 1 above, the first
The means is a process of dividing the spectral information of each band by a value corresponding to the sum of the spectral values in each band of the short-time spectral pattern, the second means is a logarithmic transformation process, and the third means
A speech recognition device characterized in that the means is a process of subtracting a value based on the average value of the spectral values of each band of the spectral pattern from the spectral value of each band.