JP2574242B2 - Voice input device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice input device that copes with fluctuations in voice speed.

2. Description of the Related Art Conventional voice input devices of this type include, for example, Takeo Watanabe, segmentation-free syllable recognition in continuous three-spoken Japanese (“Segmen
tation−free Syllable Recognition In Continuously
Spoken Japanese "), ICS SP (ICAS
SP) -83, pp. 320-323, 1983.
The configuration was as shown in the figure.

That is, a voice input terminal 31, a feature extraction unit 3 for converting an input voice signal into an input pattern comprising a sequence of feature vectors
2.Syllable standard pattern storage unit 3 that stores syllable standard patterns
3.A vowel part identification unit 34 for detecting and identifying a vowel part of an input pattern, a pattern matching unit 35 for calculating a pattern distance between a partial pattern of the input pattern and the syllable standard pattern while expanding and contracting on the time axis using DP matching. A syllable string determining unit 36 that determines a syllable string pattern of the input voice by determining a syllable standard pattern string that gives the minimum value of the cumulative distance of the inter-pattern distance obtained by the pattern matching unit 35, and a recognition result output terminal 37. The vowel part of the input speech is identified, and the pattern distance between a partial pattern for each vowel part and a syllable standard pattern having the same vowel part as the result of the identification of the vowel part is obtained. The input speech is recognized by determining a syllable standard pattern sequence that minimizes the syllabic sequence and determining the syllable sequence of the input speech.

Problems to be Solved by the Invention However, when recognizing input speech using the speech recognition device having such a configuration, although the time axis is expanded and contracted by DP matching, the utterance speed of the input speech is extremely slow. If it is faster or faster, there is a problem that extra vowel parts are added in the identification of vowel parts, or vowel parts cannot be detected and fall off, thereby deteriorating recognition accuracy.

Thus, according to the present invention, when the utterance speed of the input voice is extremely slow or fast, the speaker can ask the recognizer to say, "Speak a little faster" or "Speak a little more slowly." ”To minimize fluctuations in the speaker's voice speed,
This prevents deterioration of recognition accuracy.

Means for Solving the Problems A technical means of the present invention for solving the above problems is to provide an utterance speed instruction unit for giving an instruction relating to the audio speed in order to control the utterance speed.

Operation The operation of this technical means is as follows.

In other words, if the utterance speed is extremely slow or fast, the utterance speed instruction unit will issue an instruction such as "Please speak a little faster" or "Please speak a little more slowly", correct the speech speed, and make a speech. Speed can be kept constant.

As a result, in the recognition device, it is possible to prevent the addition of extra syllables that are uttered because the utterance speed is extremely low, and to prevent the syllables from being dropped because of the extremely fast utterance speed. It is possible to prevent the deterioration of the recognition accuracy due to the above.

Embodiments Hereinafter, embodiments of the present invention will be described, but before that, a word speech recognition apparatus using pattern matching will be described. The general configuration of this device is as follows.

A feature extraction means for converting the input speech signal into a sequence of feature vectors by a filter bank, frequency analysis LPC analysis, and the like; and a feature vector sequence uttered in advance and extracted by the feature extraction means as a standard pattern for all recognized words. Standard pattern storage means to be registered; similarity as a series of feature vectors with all of the input patterns uttered for recognition and extracted by the feature extraction means and the standard patterns stored in the standard pattern storage means Alternatively, it comprises pattern comparison means for calculating a distance, and determination means for determining and outputting as a recognition result a word corresponding to a standard pattern having the highest similarity (smallest distance) as a result of the pattern comparison.

At this time, even if the same speaker utters the same word, the utterance time length is different each time the utterance is made. Therefore, when comparing the standard pattern and the input pattern by the pattern comparing means, the time axis of both of them is Must be expanded and contracted, and the pattern lengths of both need to be aligned for comparison. At this time, since the change of the utterance time length does not occur uniformly in each part of the utterance word, it is necessary to expand and contract each part unevenly.

FIG. 3 illustrates this in a diagram. Fig. 3 (a)
, The horizontal axis represents the input pattern A = a ₁ , a ₂ ,..., A _I ( _ai is the feature vector of the i-th frame of the input pattern).
Coordinates and vertical axis are standard patterns (R ⁿ _jn is the feature vector of the j-th frame of the standard pattern R ⁿ )
Represents the j coordinate corresponding to. Input pattern A and standard pattern R ⁿ
Is matched by expanding and contracting the time axis non-linearly. In this grid graph, the path 1 indicating the correspondence between the feature vectors of the two patterns is defined as the evaluation criterion that the distance as a series of the two patterns is minimized. The heading is used as the basis, and the distance at that time is defined as the distance between the two patterns. As a method for efficiently performing this calculation, a method using dynamic programming is well known, and is called DP matching.

When determining this route, a limiting condition is set in consideration of the nature of the sound. FIG. 3 (b) shows an example of a condition for selecting a path called an inclination limit. That is, in this example, the path from the point (i-2, i-1) to the point (i-1, j) is the path from the point (i-2, i-1) to the point (i-1, j-1). Or from the point (i-
This means that only one of the routes from (1, j-1) to the point (i, j-1) can be taken, and the condition that the input pattern and the start and end of the standard pattern must always correspond If it is used, the path of the matching is limited to the hatched portion in FIG. 3 (a). This restriction is to prevent a very extreme correspondence from occurring due to the fact that no matter how the time axis expands or contracts, the same word should not expand or contract so much.

The distance between the two series, the input vector a _i and the standard pattern vector ▲ r ⁿ _j ▼ vector distance d ⁿ (i, j) of is defined as the average weighted along the path of. At this time, the DP matching method can be used if the sum of the weights along the path is kept constant regardless of the selection of the path.

FIG. 4 illustrates a matching state between a word standard pattern and an input pattern formed by combining single syllable voice standard patterns. In the figure, R ^{q (1)} , R
^{q (2)} and R ^{q (3)} mean standard patterns of single syllables q (1), q (2) q (3), and in this example, single syllables q (1), q (2) q
A case is shown in which the standard pattern of the word consisting of (3) is matched with the input pattern. According to the above description, the matching path is, for example, as follows.

Hereinafter, an embodiment of the present invention using the above-described pattern matching method will be described.

FIG. 1 is a block diagram showing one embodiment of the present invention.
In FIG. 1, reference numeral 1 denotes an input terminal of an audio signal, and reference numeral 2 denotes a feature extraction unit configured by a filter bank or the like, which converts the input audio signal into a sequence of feature vectors. Reference numeral 3 denotes a syllable standard pattern storage unit which stores a standard pattern converted into a series of feature vectors of each syllable. Here, the speech standard pattern may be defined as only a single syllable standard pattern, or may be defined as an articulatory combination generated when a single syllable is continuously uttered (for a feature vector when a single syllable speech is uttered alone, A phenomenon in which the feature vector of an uttered monosyllable voice changes under the influence of the voices before and after the monosyllable voice), and a monosyllable standard pattern and a VCV syllable standard pattern (V: vowel,
C: consonant), but the following description defines only a single syllable standard pattern. However, when a syllable is defined as a standard single syllable pattern and a standard VCV syllable pattern, only a single syllable standard pattern is sufficient for recognition of a single syllable.
In the case of word recognition, not only a single syllable standard pattern but also VCV syllables can be used, and the problem of articulation coupling can be solved.

Reference numeral 4 denotes an inter-vector distance calculation unit, which is a vector r ⁿ _j that constitutes the standard pattern R ⁿ of the syllable standard pattern storage unit 3.
Between the vector and the vector a _i that constitutes the input pattern A
Calculate d ⁿ (i, j). Now, a _i = (a _i1 , a _i2 , ……,
a _il ), Then d ⁿ (i, j) is most simply Given by Reference numeral 5 denotes an inter-vector distance storage unit which stores a result calculated by the inter-vector distance calculation unit 4. Numeral 6 denotes a cumulative distance calculation unit. For example, in FIG. 5, in an i-th frame, n = 1, 2,..., N (N is the number of syllable standard patterns) and a speech standard pattern Each Vector ▲ r ⁿ _j ▼ the input pattern A = a _1, a _2,
..., a distance d between the vector and the vector a _i of the i-th frame of a _I
ⁿ (i, j) is read out from the inter-vector distance storage unit 5, and the combined pattern of ^{Rq (1)} , ^{Rq (2)} , and ^{Rq (3)} is read. To find the cumulative distance between vectors and a _i . FIG. 3 (b) is adopted as the constraint condition of the matching path, and if the weighting factor along each path is a numerical value given on the path in FIG. 3, the cumulative distance D with respect to the standard pattern R ^{n at the} coordinates (i, j) is obtained. ⁿ (i,
j) is given as follows.

Reference numeral 7 denotes a syllable string judging unit, which determines a syllable string giving the minimum value among the final cumulative distances as a recognition result.
Is output from the recognition result output terminal 9 via the. Reference numeral 8 denotes a mode selector, which switches between two modes, a voice input mode and a correction mode. 9 is a result output terminal. Reference numeral 10 denotes a correction unit, which is constituted by, for example, a keyboard. When a speaker finds an error in the recognition result, the mode selection unit 8 is switched to a correction mode, and the error part is corrected.
The result of the error correction is output from the recognition result output terminal 9 via the mode selection unit 8 and the instruction activation unit
Output to 13. The mode selection unit 8 emits a buzzer sound when a voice is input in the correction mode, for example, to urge the speaker to switch to the voice input mode. That is, the mode selection unit 8 is switched to the correction mode only when the recognition result needs to be corrected, and is normally in the voice input mode. When the mode selection unit 8 is switched to the correction mode, the mode selection unit 8 issues a correction mode signal to the instruction activation unit 13. 11
Is a standard voice duration storage unit, which stores a standard voice duration corresponding to the number of syllables. Reference numeral 12 denotes a speech time length calculation unit, which sets a threshold value for a time series of energy levels obtained from, for example, a feature vector sequence, first detects a beginning of the input speech, starts counting speech duration, and terminates. obtaining a voice time length t ₁ by ending the count of the voice time length in the time of detecting the. The voice time length t ₁ is sent to the instruction starting unit 13. Reference numeral 13 denotes an instruction starting unit, and when the correction mode signal is input, the instruction starting unit
10 corrected recognition result by counting the number of syllables (correction result), the reading from the standard utterance time length memory 11 standard utterance time length t ₂ which corresponds to the number of syllables of the correction result, the voice time length t The ratio α to ₁ (= t ₂ / t ₁ ) is obtained, and the threshold values _THl and TH _h (T
For H _l <TH _h ), if α <TH _l , the utterance speed indicating unit 14
Against, an indication that outputs a signal S ₁ to generate an indication that "should be a little more slowly speaking", α> if TH _h, to the utterance speed instruction unit 14, "should be speaking a bit faster." and outputs a signal S ₂ for generating. TH _l ≤
signal does not occur in the case of α ≦ TH _h. Reference numeral 14 denotes a utterance speed instruction unit, which outputs an instruction `` Please speak a little more slowly '' from an instruction output terminal 15 when the signal S ₁ is input from the instruction activation unit 13, 13
If the signal S ₂ is input from, and output from the instruction output terminal 15 an instruction "Please say a bit faster." 15 is an instruction output terminal.

As described above, according to the present embodiment, when the recognition result is corrected by the correction unit, the utterance speed instructing unit gives an instruction regarding the utterance speed to the speaker as needed, so that Control the utterance speed to keep it as constant as possible,
It is possible to prevent the addition of extra syllables that are uttered because the utterance speed is extremely low, and to prevent the syllables from being dropped because of the extremely high utterance speed, and to reduce recognition accuracy due to fluctuations in the utterance speed. That can be prevented.

In this embodiment, instructing activation unit 13, in the correction mode, when the value of the ratio α of the voice time length and the standard utterance time length is not between the threshold TH _l and TH _h, Although the utterance speed instructing unit 14 is configured to issue the utterance instruction activating signal, a vowel stationary unit that detects the number (β) of vowel stationary units from the feature vector is provided.
With respect to said correction result number of syllables obtained in 10 (gamma) and the beta value, beta <when the gamma generates the signal S _1, β>
when the γ can be configured to generate the signal S _2. In this case, the speech time length calculation unit 12 and the standard utterance time length storage unit 11 become unnecessary. However, the vowel stationary part detection unit is required.

Also, each component of the above-described embodiment can realize its function by software means.

The voice input device of the present invention is in the correction mode, and when the recognition result is corrected by the correction unit, when the standard utterance time length of the correction result and the voice time length of the input voice are extremely different, By giving the speaker an instruction regarding the utterance speed from the utterance speed instruction unit, the utterance speed of the speaker is controlled, and extra syllables generated due to the extremely low utterance speed and the utterance speed being extremely fast In this way, it is possible to prevent the syllables from falling off, and to prevent the recognition accuracy from deteriorating due to the utterance speed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention, FIG. 2 is a block diagram showing a conventional example, FIGS. 3a and 3b are explanatory diagrams of the principle of DP matching, and FIG. FIG. 4 is an explanatory diagram of a principle of recognizing an input voice using a syllable standard pattern. 2 ... feature extraction unit, 3 ... syllable standard pattern storage unit, 4 ...
... inter-vector distance calculation unit, 5 ... inter-vector distance storage unit, 6 ... cumulative distance calculation unit, 7 ... syllable string determination unit, 8 ...
… Mode selection unit, 10… correction unit, 11… standard utterance time length storage unit, 12… voice time length calculation unit, 13… instruction start unit,
14 ... Speaking speed indicator.

Claims (1)

(57) [Claims] An input speech signal is converted into a sequence of feature vectors (a ₁ , a
_{2, ..., a i, ...} , a feature extraction means for converting the input pattern A composed by a _I), a standard pattern of syllables _{^{R n = (▲ r n 1}} ▼, ▲ r
_{^{n 2 ▼, ..., ▲ r}} n j ▼, ..., ▲ r n J ▼,) (n = 1,2, ...,
And syllable reference pattern storage means for storing N), the reference pattern feature vectors ▲ r constituting _{^{R n n j ▼ (j =}} 1,2 ,,
J _n ) and the inter-vector distance calculation means for calculating the inter-vector distance d ⁿ (i, j) between the feature vector a _i of the i-th frame of the input pattern A and the inter-vector distance calculation means. An inter-vector distance storing means for storing the inter-pattern distance in association with each syllable standard pattern; and the vector between each vector constituting each of the standard pattern sequence and the input pattern expressed as a combination of the standard patterns. Cumulative distance calculating means for calculating the cumulative distance of the inter-distance d ⁿ (i, j); syllable string determining means for determining a syllable string based on the result of the cumulative distance calculating means; Correcting means for correcting, a mode selecting means for enabling switching between a voice input mode and a correcting mode, a standard uttering time length storing means for storing in advance a standard uttering time length corresponding to the number of syllables, A speech time length calculating means for detecting a beginning and an end of a voice to obtain a time length between the beginning and the end, and a "fast speech when the ratio between the speech time length and the standard sounding time length is larger than a predetermined range. Utterance speed instruction prompting "
If the ratio is smaller than the range, an utterance speed instruction unit that issues an utterance speed instruction that prompts “slow utterance”;
When the result of the syllable string determination means is corrected by the correction means in the correction mode by the mode selection means,
A standard utterance time length corresponding to the number of syllables as a result of the correction is read out from the standard utterance time length storage means, and when the ratio between the voice time length and the standard utterance time length is not within a predetermined range, an utterance speed instruction is prompted. A voice input device comprising: an instruction activation unit.