JP2891259B2 - Voice section detection device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Summary] In a speech segment detection device in a speech recognition device, a speech segment threshold value at the beginning of a speech signal is set small to reduce the probability of speech dropout, and the threshold value at the end of the speech signal is reduced. This is set to be larger than the threshold value to lower the probability of adding noise. BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice section detection apparatus, and more particularly, to a detection apparatus that facilitates discrimination between voice sections, unvoiced sections, and noise in voice recognition using a computer. [Prior Art] A method of automatically detecting a voice uttered by a human by using a computer to extract features has already been widely applied.
As one of the typical methods, segmentation is performed to divide a continuous vocalized speech signal into single syllables and phonemes, and the single syllable is subjected to speech recognition. By recognizing monosyllables, it can be extended to more advanced word recognition and conversation voice recognition. At present, there is no method capable of performing a complete segmentation, but for example, a method in which a power value of a single syllable exceeds a predetermined threshold value is regarded as speech is known. That is, when the power value exceeds the power threshold (P _L ) for a certain period of time (L _v ) or longer, the section is regarded as speech. FIGS. 5A to 5C are pattern examples showing the relationship between the power value (P) of the audio signal and the utterance time (T). Here T ₀ is the voice section. (A) is, for example, “A”, “O”,
"Also", if you say "Ri", there is no problem in recognition due to the presence of all the single-syllable power for the threshold L _v in a speech period equal to or higher than the threshold value P _L. In the case of (b), for example, "a", "have", "Chi" as when uttered, with a time threshold L _S silent section, when the following threshold L _S "A" ,
“I” and “Chi” are regarded as being made by one utterance.
Silent interval of time L _S can be regarded as a low power (threshold P _L below) speech. In the case of (c),
For example, when you say “sa”, “tsu”, “ぽ”, “ro”,
The interval of “tsu” is equal to or less than the threshold value P _L and the time threshold value
Since it is longer than L _S, it is difficult to determine whether it is voice or noise. 6 (a) to 6 (d) are pattern diagrams for explaining a conventional detection method. (A) shows that all voice segments T ₀ are threshold values
There is no problem of recognition because it is P _L or more. (B) the interval T ₁ is not time threshold L _v or less, such since speech segment regarded as noise in a speech period. (C) is the interval T ₂ and T ₃ threshold L _v
More deemed large, so the speech segment, segment T ₄ are not considered to be the threshold value L _S hereinafter, such the noise of the silent section. After all, in this case, the section (T ₂ + T ₄ + T ₃ ) is regarded as the voice section.
(D) are section T ₅ and T ₇ are considered threshold L _v or less, such the noise, the interval T ₆ are considered speech section because more threshold L _v. [Problems to be Solved by the Invention] However, there is the following problem when detected by the above method. That is, at the beginning (start end) of the audio signal, the sound is likely to drop out, and at the end (end) of the audio signal, noise is likely to be added. Such start (or start) and end (or end)
There are two factors that cause different tendencies between the two. One is
In the case of Japanese, the first syllable of a word tends to be short and the syllable at the end of the word tends to be uttered longer. Is that many appear. In the case of the latter, it can be regarded as a voice because it is the sound emitted by the speaker itself,
In the case of performing voice recognition, if this part is included in a voice section, it causes erroneous recognition, and therefore, it is not preferable to include this part in the voice section. [Means and Actions for Solving the Problems] The present invention is to provide a voice section detection device which solves the above-mentioned problems, and the principle of the present invention is to detect voice (especially word voice) in voice detection. located varying the time threshold interval in the word head and endings, specifically, reduced to set the first threshold value L _V of the speech segment in the prefix in the ending than this threshold This is to set a large second threshold value. As a result, the dropout of speech at the beginning of a word and the addition of noise at the end of the word, which have conventionally been problems, can be reduced, and the accuracy of speech section detection can be significantly improved. 1 (a) and 1 (b) are characteristic diagrams for explaining the principle of the present invention. (A) shows the case of the beginning of the voice, and (b) shows the case of the end of the voice. In (a) and (b), the vertical axis PRO
Is the probability of noise addition and the probability of voice dropout.
L _V is a time threshold of the voice section. Also, I _a and I _b
The probability curve of noise addition, II _a and II _b are probability curve of the speech falling and III _a and III _b, is an error probability curve for obtaining the optimum value of L _v. (A), a is in the case of prefix is decreasing the probability of noise addition, as shown in more I _a greater the threshold L _v, conversely, the probability of speech falling rapidly as shown in II _a To increase. Also, as the threshold value is reduced, the probability of noise addition sharply increases, and the probability of voice dropout decreases.
From these curves, the curve I _a and II _a curve III _a is the sum of the curve with a minimum value as shown. Assuming that the threshold value at this minimum value is L _va , L _va indicates an optimal threshold value at the beginning of the word, and this threshold value L _va is an effective threshold in which the probability of noise addition and the probability of speech dropout are balanced. Value. In this case, _Lva varies depending on the noise environment, etc.
s. (B) shows the case of the ending. In the case of endings L _v is closer to those of large compared to the prefix. Since the pattern is the same as (a), the detailed description is omitted, but L _vb indicates the optimum threshold value at the end of the word, and is about 125 ms. That is,
The effective value of the threshold L _vb at the end is 125 ms in which the probability of noise addition and the probability of voice dropout are balanced. As described above, it has been found that by changing the threshold value between the beginning and the end of the voice section detection, it is possible to detect both the noise addition probability and the voice dropout probability. [Embodiment] FIG. 2 is a schematic configuration diagram of an apparatus for realizing a voice section detection apparatus according to the present invention. The audio signal input from the microphone 21 is subjected to high-frequency emphasis in the pre-emphasis unit 22, and then one of the power value extraction units 23 extracts the energy distribution, which is one of the characteristic parameters of the audio, by sampling. The characteristic is extracted in a band-pass filter unit 24 including a plurality of filters. The section detection unit 26 detects a voice section based on a time series PW (i) of power values, as shown in FIG. 3 described later. The voice recognition output unit 27 has a voice dictionary, performs pattern matching while referring to the voice dictionary, and outputs a recognition result from the speaker 28. The control unit 25 controls the section detection unit 26, the voice recognition output unit 27, and the like. FIG. 3 is a block diagram showing the section detecting section 26 of FIG. 2 in detail. In FIG. 3, reference numeral 261 denotes the beginning of the voice (starting point)
, 262 an end detection unit for detecting the end (end), and 263 a threshold storage unit for storing various threshold data P _L , L _va , L _vb , L _{s and the} like. For example, a sampling value PW (i) of 10 ms of the power value is input to the start end detection unit 261 and the end detection unit 262 from the power value extraction unit 23 in the previous stage. Magnitude of the time-series PW (i) the threshold P _L and power of the power that is read from the threshold storage unit 263 for each frame in the start detection section 261 are compared, further, the first threshold of the prefix The value L _va , the threshold L _{s of the} unvoiced section, and the position of the sampling frame are compared. Termination detector 262
Power threshold P _L and chronological PW in each frame as well
The magnitude of the (i) are compared, further a second threshold L _vb endings, the position of the threshold L _s and the sampling frame unvoiced is compared. In the end detection unit 262, the start detection unit 261
, And outputs the start and end position information S to the speech recognition output unit 27. FIG. 4 is a flowchart of a process in the section detection unit in FIG. Steps 1-9 in the first half of the flowchart
Is the processing in the start end detection unit 261, and the latter half of steps 10 to 21
Is a process in the end detection unit 262. In FIG. 4, i is the sampled frame number, i _S is the start frame number of the threshold, j is the number of frames continuously exceeding the threshold on the start end side, and i _e is the threshold of the threshold. The ending frame number, k, is the number of frames continuously falling below the threshold value on the ending side. As shown in the flowchart,
Comparing the magnitude for the time series PW (i) a power value threshold P _L and each frame of the power values from the power value extracting unit 23 step if (Step 3), PW (i) <a P _L 2
Is repeated. The frame number i _S when a PW (i) ≧ P _L is stored, while the PW (i) ≧ P _L continues step 6 is repeated Kaee. In step 8, PW (i)
<When P _L is determined whether the prefix of the threshold L _va, (step 9), the number of frames j enters the next termination if exceeds the threshold value L _va. If not, it is determined that no voice has been input, and the process returns to step 2. Similar steps are taken at the end, but in step 14 the PW
(I) <If not unvoiced L _S whether the determination is made unvoiced in step 21 when the P _L, i.e., the speech section detection start detection unit if larger than the number of frames k are L _S And returns to step 12 if it is a small and unvoiced section. Then, if PW (i) ≧ P _L in step 19, it is determined whether or not ending threshold L _vb at step 20, step 21 if larger than the threshold value L _vb is the number of frames j Then, the threshold value L _S of the unvoiced section is determined, and the detection of the voice section ends. After all, the start frame of the voice is i _s and the end frame is i _e
Will be obtained as [Effects of the Invention] As described above, according to the present invention, the thresholds of the beginning and the end are changed in speech section detection, so that dropout of speech at the beginning and addition of noise at the end can be significantly reduced. Thus, the accuracy of voice section detection can be significantly improved.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a characteristic diagram for explaining the principle of the present invention, FIG. 2 is a configuration diagram of an embodiment of the present invention, FIG. FIG. 4 is a processing flowchart of the present invention, FIG. 5 is a pattern diagram showing the relationship between the power value of speech and the utterance time, and FIG. 6 is a pattern diagram explaining a conventional detection method. (Explanation of symbols) 21: microphone, 22: pre-emphasis unit, 23: power extraction unit, 24: band-pass filter unit, 25: control unit, 26: section detection unit, 27: voice recognition Output unit, 28 speaker, 261 start-end detection unit, 262 end-end detection unit, 263 threshold storage unit.

Claims (1)

(57) [Claims] A voice section detection device of a voice recognition device, comprising: a start edge detection unit for comparing a voice power value at a start edge of an input voice signal with a predetermined power threshold value and a first time threshold value of a voice interval; An end detection unit for comparing a sound power value at a terminal part of the terminal with the predetermined power threshold value and a second time threshold value larger than the first time threshold value of a sound section; And a threshold storage unit for storing the first and second time thresholds. Upon detection of a voice section, the first time threshold at the beginning of the voice signal and the threshold at the end of the voice signal A voice section detection device that compares the second time threshold value and detects a voice section.