JPS62294298A - Voice input unit - 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 3. Detailed Description of the Invention Field of Industrial Application The present invention relates to a voice input device that can cope with variations in speech rate.

2. Description of the Related Art Conventional audio input devices of this type include, for example, a 5-point
ken Japanese, “ICASSP-83゜1
111) 320-323.1983. As shown in Figure 3, the configuration was as shown in Figure 3.

That is, the audio input terminal 31. Feature extraction unit 3 converts the input audio signal into an input pattern consisting of a series of all feature vectors
2. Syllable standard pattern storage unit 3 that stores syllable standard patterns
3. A vowel part identification unit 34 that detects and identifies the vowel part of the input pattern. A pattern matching unit 35 that uses DP matching to calculate the inter-pattern distance between a partial pattern of the input pattern and the syllable type pattern while expanding and contracting the time axis. A syllable string determining section 36 that determines the syllable string of the input speech by determining the syllable standard pattern string that gives the minimum cumulative distance between patterns obtained by the pattern matching section 35, and a recognition result output terminal 37. The vowel part of the composed and input speech is identified, and the inter-pattern distance between the partial pattern for each vowel part and the syllable standard pattern that has the same vowel part as the vowel part identification result is calculated, and the cumulative distance of the inter-pattern distance is calculated. The input speech is recognized by determining the syllable standard pattern string with the minimum value and determining it as the syllable string of the input speech.

Problems to be Solved by the Invention However, when recognizing input speech using a speech recognition device with such a configuration, although the time axis is expanded and contracted by DP matching, the speaking speed of the input speech is extremely slow. If the speed is too fast, there is a problem in that when identifying vowel parts, extra vowel parts are added or vowel parts cannot be detected and are omitted, resulting in deterioration of recognition accuracy.

Therefore, in the present invention, when the speaking speed of the input voice is extremely slow or fast, the recognition device can tell the speaker, ``Please speak a little faster,'' or ``Speak a little more slowly.''
(By generating instructions such as "Please speak υ")
, to prevent deterioration in recognition accuracy by minimizing fluctuations in the speaker's speaking rate.

Means for Solving the Problems The technical means of the present invention for solving the above-mentioned problems lies in the provision of a speech rate instruction section that gives instructions regarding the speech rate in order to control the speech rate.

action The effect of this technical means is as follows.

In other words, if the speaking speed is extremely slow or fast, the speaking speed instruction unit issues instructions such as "Please speak a little faster" or "Please speak a little more slowly" to correct the speaking speed and adjust the speaking speed. can be kept constant.

As a result, in the recognition device, it is possible to prevent the addition of extra syllables that occur due to extremely slow speaking speeds, and the omission of syllables that occur due to extremely fast speaking speeds.
This makes it possible to prevent the deterioration of the recognition level 8 due to variations in speech rate.

Embodiments Below, embodiments of the present invention will be described, but first a word speech recognition device using pattern matching will be explained. The general configuration of this device is as follows.

A feature extracting means converts the input speech signal into a series of feature vectors using a filter bank, frequency analysis, LPC analysis, etc., and a feature vector series uttered in advance and extracted by the feature extracting means is used as a standard pattern for all recognized words. A standard pattern storage means to be registered, an input pattern uttered for recognition and extracted by the feature extraction means, all standard patterns stored in the standard pattern storage means, and similarity as a series of feature vectors. Or a pattern comparison means for calculating distance,
It consists of a determining means for determining and outputting a word corresponding to the standard pattern with the highest degree of similarity (smallest distance) as a recognition result as a result of pattern comparison.

At this time, even if the same speaker utters the same word, the duration of the utterance differs each time, so when comparing the standard pattern and the input pattern using the pattern comparison means, the time axis of both is It is necessary to expand and contract the pattern lengths of the two to make them the same and compare them. At this time, since the change in utterance time length does not occur uniformly in each part of the uttered word, it is necessary to expand and contract each part non-uniformly.

Figure 4 represents this graphically. In Fig. 4a, the horizontal axis represents the input cover turn A=&,, aI1,...
・, i coordinate corresponding to ai (ai is the feature vector of the first frame of the input pattern), the vertical axis is the standard butter: yR
"=r"r" -=-, r"jn (r"j is the feature vector of the first frame of the standard pattern Hn) Matching by expanding and contracting means that on this lattice graph, path 1 indicating the correspondence between the feature vectors of both patterns is found based on the evaluation criterion that the distance between the two patterns as a series is minimized, and then The distance between the two patterns is the distance between the two patterns.A method using dynamic programming is well known as a method for efficiently performing this calculation, and is called DP matching.

When determining this route, limiting conditions are set in consideration of the nature of the voice. FIG. 4 is an example of a route selection condition called slope restriction. That is, in this example, the path leading to point (i, j) is the path passing from point (i-2, 1-1) to point (i-1, k), and the path leading to point (1-1, j-1) A path from or a point (
This means that only one of the routes passing from point (i, j-1) k from point (i, j-1) k can be taken, and the condition that the starting and ending ends of the input cover pattern and the standard pattern must correspond. If , the matching path is limited to the shaded area in Figure 4B. This restriction is made to prevent extreme correspondences from occurring due to the fact that no matter how much the time axis expands or contracts, it is unlikely that the same word will expand or contract so drastically.

The distance between both series is the distance between the input vector a and the standard pattern vector r and n, J'4dn (i.

)) is defined as the weighted average along said path. At this time, the DP matching method can be used if the sum of the weights assigned to the routes is made constant regardless of how the routes are selected.

FIG. 6 illustrates the matching of word standard patterns constructed by combining monosyllabic speech standard patterns and input kata patterns.

In the same figure, Rq (+l, Rq f21,
Rsu3) means the standard pattern of monosyllable q(11゜qυl, q131, this example is monosyllable q(++, q(21
, qf51 and an input cover pattern are matched.

According to the above description, the matching path is, for example, 2.

Examples of the present invention using the pattern matching method described above will be described below.

FIG. 1 is a block diagram showing one embodiment of the present invention. In the figure, 111 is an input terminal for an audio signal, and 112 is a feature extraction unit that converts the input audio signal into a series of feature vectors, which is composed of a filter bank and the like. A syllable standard pattern storage unit 113 stores standard patterns converted into a series of feature vectors for each syllable. Here, the syllable standard pattern can be defined as only the monosyllabic standard pattern, or the articulatory combination that occurs when all monosyllables are uttered continuously (
A phenomenon in which the feature vector of a continuously uttered monosyllabic voice changes due to the influence of the voices before and after the monosyllabic voice, compared to the feature vector of a monosyllabic voice uttered alone)
Taking into consideration, the monosyllabic standard pattern and the vCv concise standard pattern (v: vowel).

C: consonant), but in the following explanation, only the monosyllabic standard pattern is defined. However, when syllables are defined as a monosyllabic standard pattern and a VCV syllable standard pattern, only the monosyllabic standard pattern is sufficient for monosyllabic recognition, but in the case of word recognition, not only the monosyllabic standard pattern but also the VC V
V Jinji can be used, and the problem of articulatory combination can be solved.

Reference numeral 114 denotes an inter-vector distance calculation unit, which calculates the inter-vector distance d (x, j)e between the vectors r and n constituting the standard pattern Rn of the syllable standard pattern storage unit 113 and the vector ai constituting the input pattern Af. do. now,'
i"(ai+*'i2""""工1"rn-(,n,
rn , , , , , rfill), then d
nJ ko 1 ko 2 (i, j) is most simply given by. Reference numeral 115 denotes an inter-vector distance storage unit which stores the results calculated by the inter-vector distance calculation unit 114. Reference numeral 116 denotes a word dictionary, in which vocabulary is prepared in advance by inputting it as a monosyllabic symbol string (for example, a one-character symbol string) using a keyboard or the like.

Further, the standard utterance time length of the vocabulary is also prepared in advance by inputting it using a keyboard or the like. Reference numeral 117 is a word cumulative distance calculation unit which calculates the already calculated distance between vectors stored in the inter-vector distance storage unit 116 according to the order of monosyllables specified in the word dictionary 116 for the word to be matched. is read out, and the point as a word (
Calculate the cumulative distance to i, j). That is, for example, the sixth
In the figure, in the first frame, Th ""112.
......, syllable standard batano R"r" for N (N is the number of syllable species and genus patterns), * bow t """ a
``2 each vector turn A with Nido:
ai, , &2. . . . Read the inter-vector distance dn(i, j) from the vector ai of the first frame of work a from the inter-vector distance storage unit 116, and calculate Rq(11,
Hq (2), Rq (31 bonds/cutoff Rq
(1) ■Rq (21, ■Rq (5ゝ=rqf+1.r
υ'r ``”'・, rq(11゜Jq(1) rq(21, rq(21, , ,,,, rq(,2
1F, rq[31, rυl, ...-HH1r♀
Find the cumulative distance between the vectors between the blind person and the a person.

If we adopt Figure 4 as the constraint condition for the matching route and the weighting coefficient along each route is the numerical value attached to the route in the figure, then the cumulative distance Dn(i, j) with respect to the standard pattern Rn at the coordinates (i, j) is j) is a word determination unit 118, which outputs the word that gives the minimum value among the Jα final cumulative distances for each word obtained by the word cumulative distance calculation unit 117 as a recognition result, and outputs the word to a recognition result output terminal 119. Output from.

The recognition result is also sent to the instruction activation unit 21, which will be described later.

Reference numeral 120 denotes a voice duration calculation unit, which sets a threshold value for the time series of energy levels obtained from the feature vector series, first detects the start of the input voice, starts counting the voice duration, and calculates the end of the voice. The audio time length is determined by ending the counting of the audio time length at the time when the audio time length is detected. Reference numeral 121 denotes an instruction activation unit which reads out the standard utterance duration corresponding to the recognition result (word) obtained from the word determination unit 118 from the word dictionary 116 and calculates the standard utterance duration obtained by the utterance duration calculation unit 120. The ratio α to the voice time length (=standard voice time length/voice time length) is calculated, and the above α is the threshold value THE
and THb(TH, (THb), if α<THE, a signal S1 generates an instruction to the speech rate instruction section 122 to "speak a little more slowly."
If α) THb, the speech rate instruction unit 122
In response to this, a signal S2 is output that generates an instruction to ``please speak a little faster.'' THE≦α≦THb
In this case, no signal is generated. Reference numeral 122 denotes a speaking speed instruction section, which outputs an instruction "Please speak a little more slowly" from the instruction output 1 child 123 when the instruction activation section 121 receives the signal S, and activates the instruction. When the signal S2 is inputted from the section 121, an instruction "Please speak a little faster" is outputted from the instruction output terminal 123. 123 is an instruction output terminal.

Next, other embodiments of the present invention will be described.

FIG. 2 shows another embodiment, in which the instruction activation unit 121 of the first embodiment uses the ratio of the standard utterance time length to the voice time length (the α), the threshold value (the TH, and the TH, ), the above α is the above T He and THb, α
(In contrast to outputting the instruction generation signal when THl or α:>THb, the instruction activation unit 2 shown in this embodiment
21, the number β of vowel constant parts obtained by the constant part detecting unit 220 based on the amount of change in the feature vector, and the recognition result obtained by the word determining unit 218 (
Regarding the number of syllables γ of a word), when β<r, the signal S1 is output, and when β>γ, the signal S2 is output. Note that the constant vowel portion is defined by the constant portion detection unit 220 as a portion in which the number of frames in which the amount of change in the feature vector is equal to or less than a threshold value continues for a predetermined number of frames or more. In addition, while the word dictionary 116 of the first embodiment stores the vocabulary and the standard utterance duration of the vocabulary, the word dictionary 216 shown in this embodiment stores only the vocabulary. . That is, in FIG. 2, 211 to 21
5 and 217 to 219, and each unit indicated by 222 to 223 operates in exactly the same manner as in the first embodiment, including a word dictionary indicated by 216, a stationary part detection unit indicated by 22o, and 221.
The instruction activation unit shown by performs the above-mentioned operation.

As described above, according to the present embodiment, the speech rate of one speaker is controlled by providing a speech rate instruction section and appropriately giving instructions regarding the speech rate to the speaker. It is possible to prevent the addition of extra syllables and the omission of syllables that occur due to extremely fast speaking speed.
This makes it possible to prevent deterioration in recognition accuracy due to variations in speech rate.

Note that the functions of each component of the embodiment described above can also be realized by software means.

Effects of the Invention The voice input device of the present invention controls the speaking speed of the speaker by providing a speaking speed instruction section and appropriately giving instructions regarding the speaking speed to the speaker. It is possible to prevent the addition of extra syllables and the omission of syllables that occur due to extremely high speaking speeds, and it is possible to prevent deterioration in recognition accuracy due to fluctuations in speaking speed.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention, FIG. 2 is a block diagram showing another embodiment of the present invention, FIG. 3 is a block diagram showing a conventional example, and FIG. 4 aIb is a block diagram of DP matching. FIG. 5 is an explanatory diagram of the principle of recognizing word sounds using syllable standard patterns in the embodiment of the present invention. 112.212... Feature extraction unit, 113, 21
3...Syllable standard pattern storage section, 114, 214
・・・・・・Vector distance calculation unit, 116, 215・
... Hector distance storage section, 116,216...
...Word dictionary, 117.217...Word cumulative distance calculation unit, 118°218...Word determination unit, 120...Speech duration calculation unit, 121,221
...Instruction activation section, 122°222...
Speech rate instruction section, 220... Steady part detection section. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 3
Figure Ij-Jj-2) Figure 5

Claims (2)

[Claims] (1) The input audio signal is a series of feature vectors (a_1, a
＿２、・・・・・・、a_i、・・・・・・、a_I)
a feature extraction means for converting into an input pattern A consisting of a standard pattern of syllables R^n=(r^n_1, r^n_2, .
..., r^n_j, ......, r^n_J)
(n = 1, 2, ..., N); a word dictionary that stores vocabulary and standard utterance duration of the vocabulary; and the standard pattern R^n. Feature vector r^n_j (j=1, 2,..., J_
n) and the feature vector a_i of the i-th frame of the input pattern A, d^n(i, j)
, an inter-vector distance calculating means for calculating the inter-vector distance calculating means, an inter-vector distance storing means for storing the inter-pattern distance obtained by the inter-vector distance calculating means in correspondence with each syllable standard pattern, and a combination of the standard patterns. word cumulative distance calculation means for calculating the cumulative distance of the inter-vector distance d^n(i, j) between each vector constituting the word standard pattern and the input pattern; and the result of the word cumulative distance calculation means. word determining means for determining vocabulary from the word dictionary based on the word dictionary; speech duration calculating means for detecting the beginning and end of input speech and calculating the time length between the beginning and end; and issuing an instruction regarding the speaking rate. utterance speed instructing means; and instruction activating means for prompting a utterance instruction when the ratio of the standard speech duration stored in the word dictionary and the speech duration is not within a predetermined range regarding the result of the word determination means. A voice input device comprising: (2) The input audio signal is a series of feature vectors (a_1, a
＿２、・・・・・・、a_i、・・・・・・、a_I)
a feature extraction means for converting into an input pattern A consisting of a standard pattern of syllables R^n=(r^n_1, r^n_2, .
..., r^n_j, ......, r^n_J)
syllable standard pattern storage means for storing (n=1, 2, ..., N), and a feature vector r^n_j (j = 1, 2, ...) constituting the standard pattern R^n; ..., J
_n) and the feature vector a_i of the i-th frame of the input pattern A, the inter-vector distance d^n(i, j
); inter-vector distance storage means for storing the inter-pattern distance obtained by the inter-vector distance calculating means in correspondence with each syllable standard pattern; word cumulative distance calculation means for calculating the cumulative distance of the inter-vector distance d^n(i, j) between each vector constituting the word standard pattern and the input pattern; a word determining means for determining a vocabulary from the word dictionary based on the result; a speech rate instruction means for issuing an instruction regarding the speech rate; a word dictionary for storing the vocabulary; and a stationary device for detecting the number of stationary parts of the input speech. an instruction to prompt the speech rate instructing means to issue an instruction when the number of vowels constituting the word obtained by the word determining means and the number of steady parts obtained by the steady part detecting means are not equal; A voice input device comprising: activation means.