JPS63123100A - Voice recognition - 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 Field of the Invention The present invention relates to a voice recognition method that allows a machine to recognize a human voice.

Conventional technical word speech recognition devices have become increasingly popular in recent years as their capabilities have increased and they have become smaller and cheaper. However, it is not an easy-to-use device because of various usage restrictions, such as a limit on the number of words that can be used, restrictions on the environment in which it can be used, and the prohibition of saying unnecessary words. Particularly, in devices for non-specific speakers, people who are not aware of these usage restrictions end up using the devices. For this reason, it is common to use a word speech recognition device for non-specific speakers in combination with a voice response device to create a question-and-answer type system configuration. That is, by guiding the user's utterances with the guide voice emitted from the voice response device, consideration is given to allowing the user to use the device without being aware of any restrictions on its use.

For example, when reserving a train ticket using a device that recognizes 12 words such as numerals, "hai", and "iiie", the process is as follows.

Example 1: Equipment: How many tickets do you need? Please answer in numbers.

User: 4 (Yon) Device: Four Shinkansen tickets.

User: Yes.

FIG. 4 is a functional block diagram (conventional example) of a currently used recognition response system for non-specific speakers. When the task of the host computer 51 is activated, a guide voice is output from the voice synthesis section 52. When the user speaks according to the instructions of the guide voice, it is recognized by the voice recognition unit 50,
The recognition result is sent to the host computer 51. The host computer 51 creates a response sentence based on the user's intention, and outputs the response voice from the speech synthesis unit 52. The user then performs the task by making the next utterance, and so on.

The speech recognition section 50 performs recognition processing as follows. First, the acoustic analysis section 53 performs filter analysis, LPG analysis, power calculation, etc., and extracts parameters.

The speech section detection unit 54 detects the speech section of the input speech using the acoustic analysis results. In the similarity calculation unit 55,
Parameters included in the voice section and standard pattern section 5
The degree of similarity of each voice with the standard pattern stored in the table 6 is calculated, and the word corresponding to the standard pattern with the highest degree of similarity is taken as the recognition result.

Problems to be Solved by the Invention If used in a question-and-answer format as described above, anyone who has some understanding of speech recognition devices can use it.

However, people who are completely unaware of the limitations of speech recognition devices may not necessarily respond with the expected expressions.

For example, in the case of Example 1 above, the user may respond as shown below.

Example 2: Equipment: How many tickets do you need? Please answer in numbers.

Device: Shinkansen ticket 4 times.

■User: Yes, that's right.

The user's responses ■ and ■ in Example 2 are expressions that are naturally used in conversation, but in speech recognition targeting word sounds,
Since this kind of sentence speech cannot be recognized, it is often rejected or misrecognized. In other words, in the conventional word speech recognition device,
As in Example 1, it is necessary to respond using a fixed expression, and even the slightest variation in expression is not allowed.

This is a very large restriction in use, and is a major cause of narrowing the field of use of the speech recognition device.

The present invention aims to solve the problems of the conventional example and to ease the restrictions on the use of word speech recognition devices.

In other words, the purpose is to alleviate restrictions on the use of the device by providing a means to correctly recognize even when there are some extra parts in the input speech expression other than the expression to be recognized. shall be.

Means for Solving the Problems The present invention achieves the above object, and its technical means is to analyze an input signal containing noise and unnecessary sounds, convert it into an input parameter time series, and convert the input parameter time series into an input parameter time series. The similarity calculation between the subintervals of the series and the standard pattern of recognition target words created in advance is performed continuously by shifting the subintervals from the first 7 henos of the input parameter time series to the end, unit by unit interval. , one or more recognition word candidate sections with a high degree of similarity are cut out, and the word name corresponding to that section is used as a recognition word candidate.Meanwhile, the pitch frequency and its temporal change pattern are separately obtained from the input signal continuously. The speech recognition method is characterized in that a recognition word is determined from among the recognition word candidates using the pitch frequency magnitude and its change pattern in the recognition word candidate section or its vicinity.

Function: The present invention does not regard the entire speech interval as one word, but instead considers the entire speech interval as a sufficiently wide interval that includes the word to be recognized (in addition to the word to be recognized, it also includes eight extra words before and after the word and noise before and after the voice). from,
Only the words to be recognized are extracted and recognized using the similarity value obtained by pattern matching, the size of the pitch frequency, and the movement. In other words, the degree of similarity with each standard pattern of the word to be recognized is calculated for a part of a sufficiently wide interval including the word to be recognized, and then the similarity is calculated in the same way by shifting unit intervals. . . . The movement of the degree of similarity with each standard pattern is determined for the entire area. Then, a section (or sections) having a high degree of similarity is determined. Next, the pitch frequency of the section with a high degree of similarity is referred to, and only when the pitch frequency consistently shows a higher value than the surrounding areas of the section, that section is recognized as the section of the word to be recognized.

If the pitch frequency is not consistently higher than the surrounding area, that section is rejected. In this way, by using the recognition method by spotting only the part of the word to be recognized, it is possible to obtain correct recognition results even for input that contains unnecessary words or noise, and the word speech recognition system It becomes possible to ease the restrictions on the use of voice recognition and response systems and realize an easy-to-use voice recognition response system. Furthermore, this allows even people who are not familiar with the device to use it, thereby expanding the uses of the voice recognition device.

Example Examples of the present invention will be described below.

The present invention is uttered by pattern matching.

This is a method of recognizing keywords by extracting keywords from the recorded speech (word spotting) and verifying them using pitch frequency change patterns. Word spotting is a method that calculates similarities while scanning each word standard pattern over the entire input speech, and extracts the section where the similarity becomes large and the word name at that time. However, since the word extracted with the highest degree of similarity is not necessarily the correct answer, a certain criterion is set and a plurality of candidates satisfying the criterion are extracted. Then, from among these candidates, one word is narrowed down using the change pattern of pitch frequency and the similarity value.

The purpose of using pitch frequency is to take advantage of the fact that humans vocalize important parts of speech more clearly and at a higher recitation frequency than other parts. This tendency is particularly strong in conversations, based on our experience with speech analysis. In a question-and-answer sentence like Example 2, the important word (keyword) is the recognition target word, so by using the pitch frequency, it is possible to determine the position where the recognition target word exists.

To 10-7 FIG. 1 is a functional block diagram of a recognition response system using a voice recognition method in an embodiment of the present invention based on this idea. In the figure, since the functions of the host computer 51 and the speech synthesis section 520 are exactly the same as in the conventional example, only the contents of the speech recognition section 6 will be explained.

The input audio is analyzed by the acoustic analysis unit 1 into analysis sections (frames).
The LPG is analyzed for each frame, and the LPG
Cepstral coefficients are extracted.

Sampling frequency is 8Hz, frame period is 10m
A Hamming window with a window length of 20 n5ec is used for the 5ec1 analysis. The order of LPG analysis is 10th order, and the LPG cepstrum uses coefficients of 5th order (01 to Cs) and a power term CO. The pitch extraction unit 4 determines the pitch frequency for each frame and accumulates the value for a certain period of time. There are various methods for pitch extraction, but the simplest one is the waveform correlation method. If the human input signal is xl, the correlation function nτ is 11 l\-/ If τ is the maximum value of Vτ, then the pitch frequency is found as fo=8000/4 (2).

The similarity calculation section 2 successively compares the input parameters (LPC cepstral coefficients) with the standard pattern of each word in the word standard pattern section 3, and cuts out and stores the parts with high similarity as words.

The similarity calculation is performed not only in areas where speech exists, but also in a sufficiently wide range including noise segments before and after the sound, thereby eliminating the need to detect speech segments.

Next, we will explain how to spot words from noise or speech by calculating similarity.

The distance measure (statistical distance measure) used in mass and pattern matching will be explained.

If the input word audio length is linearly expanded or contracted to a constant length of J frames, and the parameter vector per frame is defined as a parameter vector, the input vector A will be as follows. However, t represents transposition.

/At-(&t, Mt, ・=-, at) where each 1
j is a p-dimensional vector.

As a standard pattern of word ωn (n=1.2...,N), the mean value vector is Il n N covariance matrix \W
Assuming n, the word with the maximum posterior probability P(ωnl/A) may be the recognition result.

From Bayes theorem, P(ωn1/A)−P(ωn)・P(AIωn)/P(
/A) (3) P(ωn), the first term on the right-hand side, can be regarded as a constant. Assuming a normal distribution, the second term is P(IAlωn)−(2π)]\Wi1−exp(
-1/2 (/A-7ttn)tW n (A#/
) n )) (4) The denominator term P(/A) can be regarded as a constant number if the input parameters are the same, but it does not become a constant when comparing different inputs. Here, P
Assume that (/A ) follows a normal distribution with mean jln z covariance matrix IW n.

P(/A)-(2π)1\Wmu1 °exp(-1/2(/A-Ila)ttW a(/A
-Ila)) (5)13 If we take the logarithm of Höhn (1), omit the constant term, and set it as Mn, we get Mn=(/A-//1n)t%w n(A, ,yln)
-(A-771a)'IW a(A-#a)+log
1lWn ll-1o llwa l
(6) Here, lWn and lWa are all set as common)
Let it be W.

That is, IW-(tw, 10W2 +...+IW n
+IW a ) / (N +1) (7) When formula (4) is expanded, Mn = (pn −IE n −/A
(8) However, IEn=2(W −/lln −IW ・//la
) (9) Cn-pn'lW'71n-Pa
-IW-/lla (10) Equation (8) is a primary discriminant with a small amount of calculation. Here, equation (8) is transformed as follows.

That is, Mn is the partial similarity for each frame d J =
@tj Ha (calculated by 8 additions and 1 subtraction of J.

Next, a method for recognizing speech by spo-sote 14 pages and a method for reducing the amount of calculation will be described using the above-mentioned distance measure.

In word spotting, a reference point i is set within a sufficiently long section that definitely includes the speech to be recognized.
Consider various subintervals with i as a reference, calculate the degree of similarity with each word for each subinterval using equation (11), and set the word with the highest degree of similarity throughout all subintervals as reference point 1.
This may be the recognition result for. Then, the same operation can be performed by advancing 1 by unit interval in the range of 1 to 1. In this example, words whose similarity is in the top three,
I'm looking for that interval.

If this similarity calculation is performed as is, the amount of calculation will be enormous, but by considering the duration of the word and limiting the subinterval length,
Further, by commonly using the partial similarity d1 during the calculation, the amount of calculation can be significantly reduced. FIG. 2 is an explanatory diagram of the method.

When matching input and word n, the subinterval length is (41<
4 < 42) to the standard pattern length J, and calculate the similarity with the end fixed for each frame.

It shows how it goes. The degree of similarity is calculated using equation (11) along a route starting from point T on QR and ending at P.

Therefore, all similarity calculations per frame are ΔP
This will be done within QR. By the way, since aIJ in equation (11) is the j-th frame component after expanding and contracting the section length 4, a corresponding input frame i' exists. Therefore, using the input vector, dj can be expressed as follows.

d(i', J)=@tj−j+i
(12) However, i' −i −r k (J) +
1 (13) Here, rk(j) is the word length and J
It is a function that relates the linear expansion and contraction of . Therefore, if the partial similarity between each input frame and εj is determined in advance, equation (11) can be easily calculated by selecting and adding the partial similarities having the relationship i''.By the way, Since ΔPQR moves to the right every frame, p
By calculating the partial similarity between [stock] j and a′1i on s, storing it in memory by the amount corresponding to ΔPQS, and shifting it for each frame, the required similarity can be calculated. are all in memory, so the calculation for determining partial similarity can be largely omitted, resulting in a very small amount of calculation.

The judgment unit 5 applies the pitch frequency change pattern (temporally smoothed) extracted by the pitch extraction unit 4 to the intervals (three in this example) extracted by the similarity calculation unit 2. to narrow down the number of recognized words to one. In other words, in the cut-out section, (1) If the section includes the part where the pitch frequency is the highest throughout the entire section, the word corresponding to that section is taken as the recognition result.

(2) If there is no section corresponding to (1), there is a section in which the pitch frequency movement is convex, and 6
The pitch frequency is sufficiently high (the peak of the convex mountain is the second)
If so, the word corresponding to that section is taken as the recognition result.

(3) When there are multiple sections that fall under (1) or (2), the one with higher similarity is given priority. If there is no applicable section, it will be treated as rejected.

Next, the above explanation will be illustrated with a concrete example using FIG. In FIG. 3, (a) shows the content of the utterance in relation to time, and uses ■ "Um, 4 pieces" from Example 2. (b) and (C) show the temporal movement of power and pitch frequency, respectively. In the thick line 31, the pitch frequency forms a peak, and at the peak, the pitch frequency reaches its maximum value throughout the entire area. (d) shows the temporal movement of the similarity with respect to each standard pattern, and the thick lines 32 to 34 are the sections cut out by the similarity calculation unit 2 (second
According to the method explained in the figure, a portion with a high degree of similarity is located at the end of the interval, so it is also expressed as such in FIG. 3).

The above two conditions (1) and (2) are applied to each of these three sections. First, 34 is rejected because it does not have the maximum pitch frequency and its motion is not convex. Both pitches of 32 and 33 are convex, but
Since the pitch frequency is maximum in the section 33, the word "yoshi" corresponding to 33 is set as the recognition result. In this way, correct recognition results can be obtained from unnecessary words and ambient noise.

In this way, by using the method of this embodiment, it is possible to correctly spot a target word from noise or simple sentences. Furthermore, since there is no need to detect voice sections in advance, the processing is simple. Furthermore, since the amount of calculation required for calculating the similarity is small, it is really easy to implement it in hardware.

Effects of the Invention In short, the present invention extracts recognition word candidates and their sections from input speech using a continuous linear expansion/contraction matching method, and also obtains the pitch frequency and its change pattern, and calculates the pitch frequency within the recognition word candidate section. When it is sufficiently high and the change pattern is convex, it identifies the recognized word, and conventional devices must be used in a very quiet environment, and the user must not speak other than the necessary words. This has the advantage of easing restrictions on usage and contributing to the expansion and popularization of word speech recognition devices.

[Brief explanation of the drawing]

FIG. 1 is a functional block diagram of a recognition and response system that embodies the speech recognition method in an embodiment of the present invention.
Figure 2 is a conceptual diagram explaining the method of calculating the similarity between input speech and word standard patterns in this embodiment, and Figure 3 (,)
-(d) are conceptual diagrams illustrating the recognition method of this embodiment using a specific example, and FIG. 4 is a block diagram of a recognition response system using a conventional voice recognition method. 1... Acoustic analysis section, 2... Similarity calculation section, 3... Word standard pattern section, 4...
- Pitch extraction section, 5... Judgment section. Name of agent: Patent attorney Toshio Nakao and 1 other person 2nd
Figure 3 of Zujiru Nu [1 □ □Speak

Claims (3)

[Claims] (1) Analyze the input signal containing noise and unnecessary speech, convert it into an input parameter time series, and calculate the similarity between a partial interval of the input parameter time series and a standard pattern of recognition target words created in advance. , continuously shift the subintervals from the beginning to the end of the input parameter time series by unit interval, cut out one or more recognition word candidate intervals with a high degree of similarity, and recognize the word name corresponding to that interval. On the other hand, the pitch frequency and its temporal change pattern are separately obtained from the input signal continuously, and the pitch frequency and its change pattern in the recognition word candidate section or its vicinity are used to determine the recognition word. A speech recognition method characterized by determining recognition words from among candidates. (2) As a method for calculating the similarity between the input parameter time series and the standard pattern, the time length of the input parameter is linearly expanded or contracted to the time length of the standard pattern, and the calculation is performed using a statistical distance measure that is converted to a posteriori probability. A speech recognition method according to claim 1, characterized in that: (3) As a method for determining a recognition word from recognition word candidates, in the recognition word candidate section, the pitch frequency stably takes the maximum value or a value similar to it in the entire input signal section, or the recognition word In the candidate section,
2. The speech recognition method according to claim 1, wherein when the pitch frequency change pattern becomes convex, a word corresponding to the recognized word candidate section is set as the recognized word.