JPS6310437B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a method for detecting the beginning of a speech word. Specifically, for example, after cutting out a speech section based on the power value of an input signal pattern sequence and performing phoneme classification based on the pattern on the series of speech sections, the beginning of a speech is detected based on the arrangement of the phoneme sequence of the speech section. By doing this, noise that occurs at the beginning of speech words (external noise, noise caused by lips, teeth, tongue, and saliva) is removed, and the voiceless consonants at the beginning of words are secured, improving the accuracy of detecting the beginning of speech words. The purpose is to improve the recognition rate of recognition devices.

It is now assumed that the input signal pattern sequence is expressed as a sequence of feature vectors as shown in equation (1).

X ₁ , X ₂ , ..., X _N ... (1) Each X _i , i=1, ... _, N is _an m-dimensional vector, _in ). Here, the feature vector is
For example, the output of an m-channel bandpass filter
It can be considered that x ₁ (t), ..., x _j (t), ..., x _n (t) are time-sampled.

Furthermore, a signal interval (time sampling interval) represented by a feature vector is expressed as a frame.

The power value series of the signal pattern series is a series of power values of each frame expressed by equation (2), and PW ₁ , PW ₂ , ..., PW _N ... (2) The signal expressed by equation (1) If the pattern sequence is obtained by a filter bank, the power value of vector X _i is PW _i = ( _n 〓 ^j=1 x _ij ² ) ^1/2 or PW _i = _n 〓 ^j=1 | x _ij | ...(3) is defined as.

However, X _i =(x _i1 , ..., x _in ).

Most of the conventional word beginning detection methods for audio signal pattern sequences are based only on the power value of the input signal pattern sequence as shown in equation (3). The power value PW of the series is the minimum power value of the audio signal for the first time.
PVM _i A frame with N or more has a constant frame length L
The case where 1 or more consecutive frames are input as the beginning of a word, and the case where the power value PW is less than or equal to the minimum power value PVM _i N of the audio signal for the first time in the series after the beginning of the word is considered as the end of the word. Voice section V of the signal pattern series was detected.

Furthermore, as shown in Fig. 2, there is a system whose purpose is to secure unvoiced consonants at the beginning of speech words, in which the segmentation of speech sections is performed using two power value levels Q ₁ and Q ₂ . A case where the power value PW of the signal pattern series is equal to or higher than the threshold _Q2 for the first time continues for a certain frame length L1' or more is considered a word-initial candidate, and the power value PW is set immediately before the word-initial candidate.
The point at which Q intersects with the threshold Q ₁ was taken as the beginning of the word.

In such a word beginning detection method based only on the power level of the signal pattern sequence and the length of the generation interval, it is difficult to detect noise that occurs at the time of speech production (such as external noise, vibrations, teeth, etc. accompanying the preparation for speech production).
In some cases, sounds such as lips and tongues colliding or noises caused by saliva may be detected as the beginning of a word.In order to remove such noise, increasing the power threshold level or setting a longer generation interval length may detect silence at the beginning of a word. It becomes impossible to secure consonants, and the accuracy of detecting the beginning of a word becomes extremely low.

For this reason, conventional speech recognition devices have been unable to accurately cut out speech sections, resulting in low recognition rates.

As described above, if the beginning of a speech pattern series is detected based only on the power value of the signal pattern series, it becomes difficult to distinguish between noise and unvoiced consonants at the time the speech is uttered, and the detection accuracy decreases. Therefore, if the signal pattern series expressed by equation (1) is obtained by a filter bank, it is conceivable to use frequency information based on the pattern of each frame.

The present invention focuses on this point, and will be described below along with examples thereof.

In addition to the power value series shown in equation (2), the low-frequency bias value series shown in equation (4), PL ₁ , PL ₂ , ......, PL _N (4), and the equation (5) shown in Assume that the power bias value series PD ₁ , PD ₂ , ......, PD _N (5) is used. Then, if the filter bank is composed of m channels and the center frequency w _co is w _c1 < w _c2 < ...... < w _cj < ...... < w _cn , then the low frequency of the feature vector X _i The bias value is defined as PL _i = _k 〓 ^j=1 | x _ij | (6) k<m/2, and the power bias value of the feature vector X is the power value PW _i defined by equation (3). PD _i =j such that min {j| _j 〓 ^j=1 |x _ij |>PW _i /2}.

In other words, in addition to the power value shown in equation (3), the voice pattern sequence is calculated based on three parameters: the low frequency bias value shown in equation (6) as frequency information, and the power bias value shown in equation (7). It is assumed that word beginning detection is performed.

Rough phoneme classification of each frame of the speech pattern series can be performed using three parameters: the power value, the low frequency bias value, and the power bias value. Here, the rough phonological classification refers to classification into voiced sounds, voiceless consonants, and silent sounds. Voiced sounds are vowels (|a|, |i|, |u|, |e|, |o|), and voiced consonants (|m|, |n|, |b|, |g|, |d|, | r
|, |Z|), semivowels (|z|, |w|) and pellicles (|x|, un), and voiceless consonants are (|c
|, |s|, |h|, |p|, |t|, |k|) and consonants (|Q|, that is, sounds), and silence is a state in which no phoneme is uttered.

FIG. 4 shows the correspondence between the rough phoneme classification and the power value PW, power bias value PD, and low frequency bias value PL in each frame of the speech pattern series. In the figure, a 20-channel filter bank with the center frequency and bandwidth shown in Figure 3 is used, and the low frequency bias value is
PL is obtained by setting K=3 (sum of three low-frequency channels) in equation (6). In the same figure, a is the low frequency bias value.
The rough phonological classification is shown when PL≦0.05×power value PW, and b of the same figure shows the low frequency bias value PL>0.05×power value.
The case of PW is shown.

Therefore, when detecting the beginning of a speech pattern series, we first cut out a speech section based on the power value of the signal pattern series, and next we extract each frame of the signal pattern series in that speech section using the power value PW _i and the low frequency bias value.
Rough phoneme classification is performed based on PL _i and power bias value PD _i , and finally based on the arrangement of phoneme sequences. When the beginning of a word is detected as described above, it is possible to detect the beginning of a word with higher accuracy based on the characteristics of the phoneme sequence of Japanese speech, the frequency of the noise at the time of utterance, and the characteristics of the interval of occurrence.

A characteristic of the phonetic sequence of Japanese speech is that syllables are composed of vowels, consonants + vowels, and consonants + semi-vowels + vowels, and consonants do not stand alone. Furthermore, the characteristic of noise at the time of speech generation is that it is a pulsed noise, so the generation interval is isolated. )it is conceivable that. In other words, the voiceless consonant at the beginning of a word is
For the first time in the phonetic series, voiced consonants of a certain length or longer are detected in a continuous unvoiced consonant interval before a continuous interval (vowel), and the noise at the time of speech generation is
The continuous voiced consonant section is detected as an isolated unvoiced consonant section (including some voiced sounds) of a certain length or less.

A method for detecting the beginning of a word in a phoneme sequence of a speech interval will be explained with reference to FIG. The figure shows the arrangement of phoneme sequences, where H indicates the starting frame of speech segment extraction, ■ indicates a voiced frame, □/ indicates a voiceless consonant frame, and □ indicates a silent frame.

First, starting after the starting frame H, a continuous section of voiced frames of a certain length L3 or more is detected, and its starting frame ip is detected (the vowel in the syllable,
detection of semivowels and voiced consonants). If there is no silent frame between frame H and frame ip as shown in FIG. 5a, frame H is taken as the beginning of the word WH (there is no isolated frame that becomes noise). If a silent frame exists between frame H and frame ip, the frame immediately after the silent frame closest to frame ip is set as the word-initial candidate frame WH1 (a silent consonant immediately before a vowel or semi-vowel is secured). As shown in Figure 5b, if there is no isolated non-silent frame (voiced sound or unvoiced consonant frame) of a certain length L4 or more between frame H and frame WH1, frame WH1 is set as the beginning of the word WH (of the voice). noise removal at the time of utterance). If there is an isolated non-silent frame of a certain length L4 or more between frame H and frame WH1 as shown in Figure 5c, frame WH1
The first frame of the non-silent frame most adjacent to is set as the word-initial WH (to ensure word-initial voiceless consonants and voiced consonants).

FIG. 6 shows a specific configuration of an apparatus for realizing the word beginning detection method of the present invention. In the figure, the input section 1 consists of a filter bank 13 and a sampler 14, and the parameter calculation section 2 consists of a power value calculator 15, a low frequency bias value calculator 16, a power bias value calculator 17, and a voice interval The extraction unit 3 includes a power value discriminator 18 and a power value series counter 19.
The phoneme classification section 4 is composed of a phoneme major classifier 20, a phoneme sequence counter 21, a phoneme register A22, a phoneme register B23, and an output gate 25. 1
2 is a microphone, 24 is a phoneme detection section, and 26 is an output terminal.

Next, the operation will be explained. The input audio signal input from the microphone 12 is passed through the filter bank 13.
and is inputted to the parameter calculation unit 2 as a signal parameter series via the sampler 14. In the parameter calculating section 2, the power value of the pattern sequence is calculated by the power value calculator 15 and inputted to the voice section extraction section, the low frequency bias value calculator 16, and the power bias value calculator 17. When the voice section signal _e1 is output from the voice section extraction section, the low frequency bias value calculator 16 and the power bias value calculator 17 calculate the low frequency bias value and the power bias value from the pattern and the power value, and calculate the phoneme. Output to the classification section. In the voice section extraction unit, a power value determiner 18 determines whether the power value is equal to or higher than a certain threshold level, and frames whose power value is equal to or higher than a certain threshold level are counted by a power value sequence counter 19, If a long frame number continues, a voice section signal e ₁ is output. The phoneme classification section 4 receives the speech section signal from the speech section detection section 3.
If e ₁ is output, the phoneme rough classifier 20 roughly calculates the
The phonemes are classified as voiced, unvoiced consonants, or silent, and are output to the phoneme sequence counter 21 and the phoneme register A22. Phonological sequence counter 21
When detecting for the first time that a voiced consonant frame continues for a certain length L3 or more, the phoneme register A22
The contents of are transferred to the phoneme register B23 in parallel. The word beginning detection unit 24 detects the beginning of a word based on the alignment of the phoneme series in the phoneme register B23, and
23 is output from the beginning of the word through the output gate 25 as the output phoneme sequence 26. When the contents of the phoneme register B23 have been output, since the phoneme register A22 has been updated successively, the contents of the phoneme register B23 are The contents of the phoneme register A22 are outputted from the output terminal 26 as an output phoneme sequence via the output gate 25 from frames that follow the contents.

As is clear from the above description, the present invention cuts out a voice section based on the power value of an input signal pattern series, and further extracts each frame of the signal pattern series of this voice section based on the low frequency bias value and power bias found from the pattern. After performing a rough phoneme classification based on the power value and the power value, the beginning of the word is detected based on the phoneme sequence of the speech interval, thereby removing noise at the time of speech production and unvoiced speech at the beginning of the speech. It is possible to secure consonants, improve the accuracy of detecting the beginning of speech, and improve the recognition rate of the speech recognition device.

[Brief explanation of the drawing]

1 and 2 are waveform diagrams showing the operation of cutting out a speech section, respectively, and FIG. 3 is a frequency characteristic diagram of a filter bank for creating a speech signal pattern series to which the speech beginning detection method according to the present invention is applied.
Figures 4a and 4b are diagrams showing the correspondence between the rough phoneme classification of each frame of the signal pattern sequence and the power value, low frequency bias value, and power bias value of the pattern. A process explanatory diagram of the operation of detecting the beginning of a speech word from the phoneme sequence of a speech interval, respectively, Part 6
The figure is a block diagram of a speech word beginning detection device to which the present invention is applied. 1...Input section, 2...Parameter calculation section, 3...
...Speech segment cutting unit, 4...Phonological classification unit.

Claims (1)

[Claims] 1. A start interval in which the power value of a series of input signal patterns that continues for a certain length is equal to or higher than the power value determined as a threshold value for the first time, and a power value for which the power value of the pattern series that continues for a certain length for the first time after that start interval is determined as a threshold value. The end section that is less than or equal to the value is detected and the speech section is cut out, and each frame of the speech section is classified into voiced sounds (vowels, voiced consonants), voiceless consonants, and silence based on the pattern of each frame. The beginning of the word in the speech section is detected from the beginning of the speech section, and if there is no silent frame in the section where voiced frames are continuous,
The starting point is taken as the beginning of a word, and if there is a silent frame, the frame immediately after the silent frame most adjacent to the continuous voiced frame section is taken as the beginning of a word,
If there is no continuous non-silent (voiced or unvoiced consonant) frame section between the starting point and the word-initial candidate, the word-initial candidate is taken as the word-initial, and if there is a continuous non-silent frame section, the word-initial A method for detecting the beginning of a word in speech, characterized in that the first frame of the continuous non-silent frame section most adjacent to a candidate is taken as the beginning of a word.