JPH0854892A - Voice recognition method - Google Patents

Abstract

PURPOSE:To execute word recognition without preparing standard patterns of unnecessary words for a voice segment which includes the unnecessary words by obtaining the degree of similarity against a sound section only and using the standard template, which makes the degree of similarity a minimum, as word recognition result. CONSTITUTION:Let the point, at which a change takes place from a silence section to a sound section in voice segment, be a starting point stq (q=1, 2...) and let the point, at which a sound section changes to a silence section, be an ending point edr (r=1, 2...). Employing the DP matching equation shown, the degree of similarity SS of the voice segment and the standard template between the point stq and the point edr is obtained. Note that in the equation, Sn (i, j) represents the DP matching score between a first frame of the inputted voice and a j-th frame of a n-th standard pattern, d (Ai, Bnj) signifies the segment distance between the first frame of the inputted voice and the j-th frame of the n-th standard pattern and a min() represents a minimum value among three scores. The word having the standard template, which makes the degree of similarity SS a minimum, is mad as the word recognition result of the voice segment.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method,
In particular, it relates to a method for recognizing a discrete word utterance.

[0002]

2. Description of the Related Art A man-machine interface for interacting with a computer allows a person to input or instruct data to the computer by using an input device such as a keyboard or a pointing device (mouse, etc.), and an output device such as a display, etc. It displays various data and responds to people.

These man-machine interfaces
In addition to the device, there is one that performs input and output by voice recently, and a computer is provided with a voice synthesizer and a voice generator as voice output devices, and a voice recognizer and a decoder as voice input devices.

As a method of recognizing a voice generated by a person, discrete word utterance sentences in which pauses (breaks) are placed between words are individually recognized, and one for decoding a sentence from a series of recognized words as a sentence. There is a Pass DP matching method. This method is an extension of the DP matching method, which is a word voice recognition method, to sentence recognition.

The DP matching method will be described below.

When performing word recognition, a speech waveform is sampled at a certain time interval, converted into a time series of multidimensional feature vectors such as a spectrum, and then handled. Similarly, the words to be recognized are converted into time series of multidimensional feature vectors, and these are registered in the computer as standard patterns.

In the voice recognition process, the similarity between the input feature vector time series and the feature vector time series of the standard pattern is calculated for all standard patterns, and the word of the most similar standard pattern is used as the recognition word.

However, it is not possible to directly compare the input feature vector time series and the standard pattern feature vector time series as they are. This is because there are individual differences in the length of time that a person utters a sentence or word, and even when the same person utters the same word, it varies greatly depending on the mood depending on the day. Moreover, the expansion and contraction of the utterance time is not uniform and varies nonlinearly.

In the DP matching method, the time axis is converted by using the dynamic programming method so that the feature vector time series of the input speech best matches the feature vector time series of the standard pattern, and then the similarity is calculated. Ask.

The concept of this DP matching will be described with reference to FIG. In the figure, the horizontal axis shows the input voice and the vertical axis shows the standard pattern of words registered in the computer. Here, it is assumed that both the input voice and the standard pattern are described in time series of phoneme labels, not in time series of feature vectors.

Normally, in DP matching, the similarity between the input voice and the standard pattern is calculated under the condition that the end points are fixed. The fixed end point means that the first frame of the input voice corresponds to the first frame of the standard pattern (fixed start point), and the last frame of the input voice corresponds to the last frame of the standard pattern (fixed end point). It is a constraint condition.

DP matching is based on this constraint condition.
The time axis is converted so that the input voice and the standard pattern match best, and the similarity between the two is obtained.

In the One Pass DP matching, the standard template of the DP matching for word recognition is changed to the one for discrete sentence recognition as shown in FIG. 6, and the end point fixed DP matching is applied to the start and end of the discrete sentence. To recognize each word contained in the sentence.

[0014]

[Problems to be Solved by the Invention] One Pass D
The P matching method has a high recognition performance in a discrete word utterance sentence containing no unnecessary words, but the recognition performance of a discrete word utterance sentence containing unnecessary words is not so high.

Here, the unnecessary word is a word that is not registered in the word standard template. For example, when a speaker originally says "Tokyo", the speaker says "er Tokyo". It corresponds to “” um ”when you do.

Originally, the One Pass DP matching is "what kind of word is included in that section to improve the recognition score by fixing the start point and the end point of the input voice?"
Since this method is a method of calculating mathmatically, when an unnecessary word is included in the utterance of a discrete word, the unnecessary word and some standard template are forcibly associated with each other, resulting in poor recognition performance.

Therefore, a method of registering unnecessary words in the word standard template in advance can be considered. However, since the unnecessary words are unpredictable utterances, it is impossible to cover all unnecessary words as a standard template. .

Further, the unnecessary words include external noise in addition to the words uttered by a person, but it is also impossible to create these standard templates.

An object of the present invention is to provide a speech recognition method which, when recognizing a discrete word utterance, eliminates unnecessary words included in utterance sounds and can recognize only original input words to be recognized. Especially.

[0020]

SUMMARY OF THE INVENTION In order to solve the above problems, the present invention provides a voice spotting process for detecting a voice section of a discrete word utterance and a feature vector time for each frame of each detected voice section. A sequence is obtained, this feature vector time series is converted into a phoneme label time series, and the similarity between this phoneme label time series and the standard template of the phoneme label time series as a dictionary is obtained by the fixed endpoint DP matching method for word recognition. In the voice recognition method by the One Pass DP matching method, which includes a word recognition process and a word determination process that narrows down the degree of similarity from a standard pattern among the recognized word candidates, the end point fixed DP matching process, The voice section from the start point at which the voice section changes from the silent section to the voice section to the end point at which the voice section changes to the voiceless section and the standard template The following equation the similarity SS of the rate,

[0021]

[Equation 2]

Sn (i, j): DP matching score between the ith frame of the input voice and the jth frame of the nth standard pattern.

D (Ai, Bnj): Partial distance between the ith frame of the input voice and the jth frame of the nth standard pattern.

Min (): Minimum value among the three scores.

According to the present invention, the word of the standard template that minimizes the similarity SS is used as the word candidate of the voice section.

[0026]

When the unnecessary section is included in the voice section because the silent section between the unnecessary word and the target word is short, the similarity by the end point fixed DP matching is calculated only for the voiced section, and this similarity is minimized. The standard template to be used is the word recognition result.

[0027]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In describing an embodiment of the present invention, it is assumed that the input voice and standard template are as follows.

(1) The input voice consists of I frames, and the i-th frame is indicated by Ai (i = 1, 2, ..., N).

(2) It is assumed that N words are registered in the word dictionary of the recognition device. That is, the number of standard templates is N
Is. The nth standard template is Tn (n = 1,
2, ..., N), Tn is composed of Jn frames, and the j-th frame is represented by Bnj (j = 1, 2, ..., N).

FIG. 1 is a flowchart showing a recognition processing of a discrete word utterance sentence according to an embodiment of the present invention. The voice spotting process S1 detects a voice section of an input voice that is discretely and continuously uttered. In the word recognition process S2, word recognition is performed on each voice section detected by the voice spotting process, and word candidates of each voice section are obtained. The word determination process S3 narrows down the word candidates from the respective word candidates based on the similarity to the standard pattern, and outputs one word candidate as a recognition result. Hereinafter, processing of each unit will be described in detail.

(1) Voice spotting process ... Each voice section is detected from a discrete word utterance using the power information of the input voice.

First, for the input speech waveform as shown in FIG. 2A, the power information is obtained by the logarithmic power of each frame by the following equation.

[0033]

(Equation 3)

P (i): Logarithmic power of i-th frame AM _K : Amplitude of k-th speech in section included in i-th frame Next, as shown in (b) of FIG. The number of frames αp (p = 1, 2,
...). When this [alpha] p is larger than the threshold value of the pause time (PAUSE), the front and rear thereof are separated as a silent section.

On the contrary, the number of frames βp (p = 1, 2, ...) In the section where the obtained logarithmic power is equal to or larger than the threshold value θ is obtained.
When this βp is larger than the threshold value of the duration (DURATION), the preceding and succeeding portions are separated as a voiced section.

By performing the above processing, the voice section (speech section) is spotted from the discrete word utterance.

(2) Word recognition process: Word recognition is performed on the spotted voice section by the procedure shown in FIG.

First, feature extraction is performed by DFT (discrete Fourier transform) for each frame to obtain a feature vector time series (S2 ₁ ).

Next, the input feature vector time series is converted into a phoneme label time series by a phoneme classifier using a neural network (S2 ₂ ).

Next, the degree of similarity with the standard template having a phoneme label time series as a dictionary is obtained by the fixed endpoint DP matching method to obtain word recognition (S2 ₃ ).

Here, when the silent section between the unnecessary word and the target word is short, both the unnecessary word and the target word are included in the voice section detected in the voice spotting process S1.

In this case, in the normal fixed-point DP matching method, since the speech segment including the unnecessary word and the target word is matched as one word, the similarity between the target word and the standard template is correctly obtained. I can't.

In order to eliminate such inconvenience, in the present embodiment, the target word is recognized by the following method.

(A) As shown in FIG. 4, the start point stq (q = 1, 2,
…), And the end point edr is the point where the sound part changes to the silence part.
(R = 1, 2, ...). And DP according to the following formula
Using matching, the similarity SS (st) between the voice segment from the start point stq to the end point edr and the standard template Tn
q, edr, n) is obtained.

[0045]

[Equation 4]

Sn (i, j): DP matching score between the ith frame of the input voice and the jth frame of the nth standard pattern.

D (Ai, Bnj): Partial distance between the i-th frame of the input voice and the j-th frame of the n-th standard pattern. The higher the similarity between Ai and Bnj, the smaller the value of the partial distance.

Min (): Minimum value among the three scores.

(B) The similarity SS (st
The standard template that minimizes q, edr, n) and the minimum value SSmin (stq, edr, n) at that time are obtained.

(C) q = 1, 2, ... And r = 1, 2 ,.
On the other hand, the standard template that minimizes the minimum value SSmin (stq, edr, n) is set as the word recognition result (word candidate) of the voice section. For this, the standard template Tn has the similarity SS (stq, edr) in the interval {stq, edr}.
edr, n).

By performing fixed-point DP matching as in the above method, it is possible to stably recognize the target word even in a voice section in which unnecessary words are not separated.

(3) Word determination process: All word candidates having a value smaller than the threshold value γ are set as final word recognition results among all the word candidates obtained by the word recognition process.

[0053]

As described above, according to the present invention, the similarity between the phoneme label time series of the voice section and the phoneme label time series of the standard pattern is obtained by the fixed endpoint DP matching, only for the voiced part. In order to obtain the similarity and use the standard template that minimizes this similarity as the word recognition result,
There is also an effect that word recognition can be performed without preparing a standard pattern of unnecessary words in a voice section including unnecessary words.

[Brief description of drawings]

FIG. 1 is a flowchart showing an embodiment of the present invention.

FIG. 2 is a diagram of a voice spotting processing mode according to the embodiment.

FIG. 3 is a word recognition processing flowchart in the embodiment.

FIG. 4 is an example of DP matching for word spotting.

FIG. 5 shows an example of conventional DP matching.

FIG. 6 shows an example of conventional One Pass DP matching.

Claims (1)

[Claims] 1. A voice spotting process for detecting a voice section of a discrete word utterance, a feature vector time series for each frame of each detected voice section is obtained, and this feature vector time series is converted into a phoneme label time series. Then, the word recognition processing for recognizing words by obtaining the similarity between this phoneme label time series and the standard template of the phoneme label time series held as a dictionary by the fixed endpoint DP matching method, and the standard pattern from among the recognized word candidates. One P with word determination processing that narrows down based on the degree of similarity with
In the voice recognition method according to the ass DP matching method, the fixed endpoint DP matching process includes a voice section from a start point of a voice section that changes from a silent section to a voice section to an end point that changes from a voice section to a silence section and the standard. The similarity SS with the template is expressed by the following equation, Sn (i, j): DP matching score between the ith frame of the input voice and the jth frame of the nth standard pattern. d (Ai, Bnj): Partial distance between the ith frame of the input voice and the jth frame of the nth standard pattern. min (): minimum value among the three scores. A speech recognition method, characterized in that the word of the standard template that minimizes the similarity SS is used as a word candidate of the speech section.