JPS62145298A - 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 that recognizes input speech information.

[Prior Art J] A speech recognition device first performs A/D conversion on input speech and sends the output to a feature extraction unit. The feature extraction section calculates voice power information and uses techniques such as fast Fourier transform to calculate voice spectral information.

The type of information held by the standard patterns stored in the standard pattern storage unit matches the type of information calculated by the feature extraction unit, and similarity calculation during pattern matching is performed based on the input audio and Calculations are made for each amount of information of the same type in a standard pattern, and the final similarity is determined by multiplying each amount of information found by a predetermined value and adding them together.

By the way, in the conventional phonetic translation:& device, voiced line and unvoiced line f
Discrimination of mute and voiceless fragments in zarananimuin,
The identification list for nasal consonants in the voice meter uses voice power information, divides the frequency range into low, middle, and high ranges, and calculates the frequency components included in that range. This was determined by using the ratio.

[Problem to be solved by the invention] However, if the input audio contains a lot of noise,
The power information of the consonant at the beginning of a word is often hidden in the power of jAg, and even for the consonant in the middle of a word, its spectrum is dragged by the spectrum of the flJ Yf before and after, and qualitative information is missing, resulting in a consonant. was not easy to identify.

Also, the spectrum of the mother = /u / is the nasal consonant /m/
, /n/, their misidentification rate was also high.

The present invention has been made in view of the above-mentioned shortcomings of the prior art, and its purpose is to shorten the processing time during matching by incorporating information about the input speech waveform into the speech recognition method, and to achieve high recognition performance. The object of the present invention is to provide a speech recognition device that can obtain a high rate of speech recognition.

L Stage for Solving Problem C] As an alternative means for solving this problem, for example, the voice recognition apparatus of this embodiment includes a first calculation means for calculating the peak level of the waveform of the 11 voice information. and a second calculation means for calculating the temporal change in the peak level obtained by the first calculation means, and a second calculation means for calculating the temporal change in the peak level of the audio information obtained by the second calculation means as a feature of the audio information. and an integration f stage for integration.

[Operations] In the configuration of this embodiment, changes in the temporal peak level of the input audio information are newly integrated into the characteristics of the audio information, and as a result, the obtained optimal recognition a result is output.

"7! Embodiments" Hereinafter, embodiments according to the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a block diagram of the speech recognition device of this embodiment.

In the figure, 1 is an audio input section that converts audio into an electrical signal, and 2 is a microphone that samples and quantizes the audio converted into an electrical signal every 5 to 10 msec. /D converter. A buffer memory 3 temporarily stores the output of the A/D converter 2. 4 sequentially reads data from the buffer memory 3 to obtain a wave height of 111'I.
4a is the central processing unit CPU in the wave height 1+11 detection i-1 circuit 4, and 4b stores the program of the flowchart shown in FIG. 3, which will be described later. 4c is a ROM, and 4c is a work area or d(, which is used in 5 when calculating the wave height adjustment described later).
This RAM is used as a buffer for 1), d(2), and d(3). 5 is a pack memory for temporarily storing the output of the wave height 11 self-detection circuit 4, and 6 is a wave height 1 self-change calculation circuit that calculates the 11-time change in the wave height value. 7 is a frequency range of 200-6000 Hz divided into 8-30 channels 1? This is a feature extraction unit that includes a group of pass-pass filters and extracts features such as power W and spectral information. 8
Until the characteristic Tit related to the wave height is calculated,
It is a buffer memory that stores the input g voice feature amount, and 9
is a feature pattern integration unit that integrates the output of the feature extraction unit and the features related to temporary +5'5 (If) to create a feature pattern based on the input il; 10 is a memory that stores standard patterns; 11 is an input block; This is a pattern matching unit that compares the J+1 feature report and the standard/41i pattern read from memory 10 and calculates the similarity between the two. 12 is the standard pattern with the maximum similarity calculated by pattern matching FAll. This is a recognition result output unit that outputs as a recognition result.

The operation of this embodiment will be explained in detail below.

First, the input voice is converted into a digital signal by the A/D converter 2, and its output is sent to the peak value detection circuit 4 and the feature extraction section 7 via the pack memory 3. Is the sampling frequency of A/D conversion and ■ the number of quantization hits per sample variable? In this example, sampling is performed at 12 kHz, and Quantized. In this case, one second of audio will be represented by 12,000 points of data.

FIG. 2 is a graphical representation of the output of the A/D converter 2.

Since A/D conversion is performed in real time, the wave height (llf
A buffer memory 3 is placed in front of the f detection circuit 4, and the wave height is
The α detection circuit 4 sequentially reads sample data from the buffer memory 3.

FIG. 3 is a flowchart of processing performed by the CPU 4a inside the peak value detection circuit 4.

Here, d(11, 6(2), d(3) is the storage 11 that stores data read from the pack memory 3, and 1
, 2.3 represents sample data 11"1. Also, d+ represents a positive wave height W, and d- represents a negative wave height 11[i.

First, the first two pieces of data stored in the buffer memory 3 in step St are imported into d(1) and d(2), respectively. In step S2, it is determined whether all the data in the buffer memory 3 has been read, and when it has been read, step S9
The process ends at step S3, but since it is not finished at this point, proceed to step S3.

Read the next data from pack memory 3 into d(3).

In steps S4 and S6, d(1), d(2), d(3
). For example, if d(1) (d(2) and d(2)) d(3), then d(2) when d(2))O
) is the positive peak value, d(1)) d(2) and d
(2) (In d(3), when d(2)co, d(2) is a negative peak value, so when each of the above conditions is satisfied, d+ and d- are set in step S5.7. d(2) is stored, and the data number ('') is stored, and the process proceeds to step S8.

In addition, if the above conditions are not satisfied, step S8 is performed directly.
Proceed to. In step S8, the current d(2) is changed to d(1)
Similarly, d(3) is stored in d(2), and the process moves to step S2, where it is determined whether or not the end is complete. If it is not the end, new data is read from the buffer memory 3, and the data is stored in d(2). (3) and repeats the same process to process all data.

In this case, the number of data is the value obtained by multiplying the measurement time by 12,000.

As mentioned above, when calculating the peak value, the following procedure is performed.

[Explanation for determining the positive peak value] In other words, if d(1)≦d(2), roughly d
Compare (2) and d(3).a(3)<d(2
), then the value of 2 is the peak, so the value of d(2) is the peak value. Furthermore, you can check whether this is a positive peak value by checking the sign of d(2), d(2)>
If it is O, d(2) is substituted for d+, the values of d+ and n are stored, and the processing from step 58 onwards is performed.

In other cases, for example, d(3)≧d(2), d
(2) In the case of ≦0, etc., the processing from step S8 onwards is directly performed.

[Explanation for finding negative wave height i1] In other words, when d(1)≧d(2), roughly d(
2) and d(3) are compared. d(3) > d(2
), the value of 2 is the peak, so the value of d(2) is the wave height (+ti).Furthermore, whether this is a negative peak value can be determined by checking the sign of d(2). d(2
) < 0, substitute d(2) for d-, and set d- and n
The value of is stored and the processing from step 88 onwards is performed.

In other cases, for example, d(3)≦d(2), d
(2) If ≧O, etc., the process from step 58 onwards is directly performed.

In FIG. 2, examples of the positions of the positive peak value d+ and the negative peak value d- determined by the peak value detection circuit 4 are represented by signals and mu, respectively.

The peak value change calculation circuit 6 of this embodiment calculates the following feature amount from the output of the peak value detection circuit 4 described above.

Furthermore, in the following equations, d"(n) and d-(n) represent a set of time information n and peak value information d+ and d-.

Ratio of the sum of positive wave height and negative wave height within a certain time; pl:Σ(d+(n); n≦Tl/Σ(d-(n);
n≦T) Ratio of adjacent wave heights 1+fi of the same sign and distance between them; p2: d+(n-1)/d+(n)p
2(n, t): (time of n for d''(n)) - (d+
(n-1) p3: I d-(n-1) I/I d
ln) Ip3(n, t): (Time of n to become d-(n)) - (Time of n-1 to become d-(n-1)) Distance between ratios of peak values of adjacent different codes ;p4(
n, +): cD(n-1)/l d-(n) 1
p4(n, t): (d+(n) at time 1(d-
(n-1) p5(n,-): I d-(n-1) I /d+
(n) p5(r+, t) : (n that becomes cD(n)
time) - (time of n-1 for d-(n-1)) In the feature pattern integration section 9, the output from the feature extraction section 7 is
A new feature pattern of the input voice is created by integrating the feature pattern stored in the buffer memory 8 and the output of the peak value change calculation circuit 6. Hereinafter, this newly created pattern will be simply referred to as a feature pattern.

In addition, in this example, sampling was performed at 12kHz, but
The feature pattern integration unit 9 sets new conditions for the feature pattern based on the following criteria.

The emblem pattern storage section 7b has been selected.

1) Discrimination between voiced and unvoiced sounds The difference in value between d+(n) and d-(n) is 100 or more, and p4(n,÷)>I,3 or P5(n,-)
) If it satisfies 0.713, it is determined to be a voiced piece.

Otherwise, it is determined that the sound is voiceless.

2) Discrimination between silence and voiceless consonants l] for those determined to be voiceless sounds.

If p2(n,t)(3 or p3(n,t)<3 is satisfied, it is determined to be silent. Otherwise, it is determined to be an abusive consonant.

3) Discrimination between vowels and consonants For those determined to be voiced in 1), Pi) 1.
If it is 5, it is determined to be a vowel. If not, it is determined to be a preliminary run.

As mentioned above, by integrating the wave height of audio information (inter-114i changes in ITi and the discrimination results of each phoneme into new audio information feature patterns in the feature pattern integration unit 9), more accurate audio information features can be obtained. It becomes possible to set.

In addition, the pattern matching unit 11 sequentially reads the standard patterns from the memory IO, calculates the degree of similarity with the feature pattern output from the feature integration unit, and outputs the standard pattern with the highest degree of similarity to the recognition result output unit 12. and output the standard pattern.

In this example, the recognition process was performed by newly integrating the time change information of the peak value as a feature parameter of the audio information.
In addition, the same effect can be obtained by using the ratio of the number of zero crossings of the audio waveform per unit time, the intensity ratio of the audio spectrum per unit time, or the like.

Further, although sampling was performed at 12 kHz in this embodiment, the frequency is not limited to this, and although one sampling is 12 bits, this is also not limited to 12 hints.

[Effects of the Invention] As described above, according to the present invention, audio special y! 1.
It has recently become possible to improve the recognition rate of speech recognition processing by adding not only peak value information but also time change information of speech to the amount.

Furthermore, since the matching candidates can be narrowed down based on the temporal change information of the audio, it is effective in shortening the processing time.

[Brief explanation of drawings]

Fig. 1 is a block diagram of the speech recognition device of this embodiment. Fig. 2 is a graphical representation of input speech output after A/D conversion. Fig. 3 is a peak value detection diagram of the embodiment. 3 is a flowchart showing processing. In the figure, 1...microphone, 2...A/D converter,
3.5.8.../heffer memory, 4... Peak value detection circuit, 4 a.-CPU, 4 b-ROM, 4
c...-RAM, 6... Peak value change calculation circuit, 7.
...Feature extraction unit, 9...Feature pattern integration unit, 10...
・Memory, 11...Pattern matching section, 12...
This is a recognition result output unit. Patent applicant Canon Co., Ltd. Figure 1 Figure 2

Claims (2)

[Claims] (1) In a speech recognition device that selects a characteristic pattern of input voice information and recognizes a standard pattern from the characteristic pattern, a first calculating means for calculating a peak level of a waveform of the voice information; a second calculating means for calculating a temporal change in the peak level obtained by the calculating means; and an integration for integrating the temporal change in the peak level of the audio information obtained by the second calculating means into a characteristic pattern of the audio information. A speech recognition device comprising: means. (2) The integrating means includes a discriminating means for discriminating the type of phoneme of the audio information from the temporal change in the peak level calculated by the second calculating means, and a discriminating means for discriminating the type of phoneme of the audio information from the temporal change in the peak level calculated by the second calculating means, and a part of the feature pattern that is determined by the discriminating means. 2. The speech recognition device according to claim 1, further comprising: a synthesis means for synthesizing the speech information.