JP3445117B2 - Acoustic analysis method for speech recognition - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an acoustic analysis method for speech recognition, and more particularly to an acoustic analysis method for speech recognition suitable for use in a speech dialogue system.

[0002]

As an example of a conventional acoustic analysis method and apparatus for speech recognition, the present invention has been invented by the present inventors.
There is a patent application (Japanese Patent Laid-Open No. 9-90990). The outline of this conventional technique will be briefly described with reference to the block diagram of FIG. M-th frame consisting of frames x1 to xT
When the th voice input M is input, the feature vector calculation unit 31 obtains a feature vector CM (t) (here, t = 1, 2, ..., T) for each frame. The feature vector CM (t) is sent to the storage unit 32 and the subtraction unit 34. The storage unit 32 includes a first storage unit 32a and a second storage unit 32b. The first storage unit 32a stores the feature vector CM (t) of the speech input M of the Mth utterance, The second storage unit 32b stores the feature vector CM-1 (t) of the voice input (M-1) of the (M-1) th utterance.

The average calculator 33 receives feature vectors CM-1 (1), CM-1 (2), ..., C of the input voice input (M-1).
From M-1 (T), the voice input (M-
Obtain the cepstrum average of 1).

[0004]

[Equation 1] The subtraction unit 34 receives the voice input M from the feature vector calculation unit 31.
Average value of the voice input (M-1) is subtracted from the feature vector CM (t), and the operation represented by the following equation (2) is performed to normalize the cepstrum average value (hereinafter abbreviated as CMN). ) Calculated cepstrum <CM (t)>.

[0005]

[Equation 2] The pattern matching unit 35 matches the CMN-processed cepstrum with the standard pattern stored in the standard pattern storage unit 36, and outputs a recognition result.

[0006]

As is apparent from the above equation (2), the CMN-processed cepstrum <CM (t)> is
It is greatly affected by the utterance before the first utterance. Therefore, this 1
When the utterance before utterance is short, for example, when the utterance before the first utterance is short such as "A" or "Yes", the cepstrum average value of the utterance sufficiently reflects the characteristics of the voice. However, if CMN is performed using this cepstrum average value, there is a problem in that the voice recognition rate becomes low.

Further, in the above-mentioned prior art, there is a problem that the voice input of only one voice cannot be recognized in real time because the voice of one voice before is required.

An object of the present invention is to eliminate the above-mentioned problems of the prior art and not to reduce the recognition rate of speech even if the utterance before one utterance is short, or for speech recognition to reduce the reduction of the recognition rate. To provide a method of acoustic analysis of. Another object of the present invention is to provide an acoustic analysis method for voice recognition, which enables real-time voice recognition without the need for voices before one utterance.

[0009]

In order to achieve the above object, the present invention obtains a feature vector from an input voice, normalizes the feature vector, and performs pattern matching with a standard pattern of voice stored in advance. In an acoustic analysis method for performing voice recognition, the average value of the feature vector and a part dependent on the utterance content estimated by using the recognition result of the pattern matching are used to cancel the part dependent on the utterance content. Estimate the cepstral mean value that does not depend on the utterance content,
The first feature is that speech recognition is performed by using the average value of the cepstrum for the normalization of the feature vector.
Further, the present invention has a second feature in that the estimation of the portion depending on the utterance content is performed by using the recognition result of the pattern matching of the utterance of one sentence before. A third feature is that the dependent portion is estimated by using the partial recognition result of the voice that is currently uttered halfway.

According to the first and second characteristics, it is possible to accurately estimate the average value of the cepstrum for a short utterance "a", for example, and even if the utterance before one utterance is short. It is possible to prevent the voice recognition rate from being reduced or prevent the reduction in the recognition rate from being reduced. Further, according to the third feature described above, it becomes possible to perform voice recognition in real time using only the voice without using the voice before one utterance.

[0011]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described in detail below with reference to the drawings. FIG. 1 shows a block diagram of an embodiment of the present invention, in which a feature vector calculation unit 1, a storage unit 2, a normalization parameter calculation unit 3, a subtraction unit 4, a pattern matching unit 5,
The standard pattern storage unit 6 and the recognition result state series storage unit 7 constitute a main part. The delay unit 1
The circuit composed of 1, the subtraction unit 12, the average calculation unit 13, and the pattern matching unit 14 is provided for recognizing only the first voice input M = 0, as will be apparent from the description below. , Which is an adjunct to the present invention. This accessory part is connected to the main part of the present embodiment by switches SW1 and SW2 which are functionally represented. The switches SW1 and SW2 are in the connection state of the solid line shown only while the first voice (M = 0) is input, and when the second voice (M = 1) is input, the switches SW1 and SW2 are represented by the dotted lines. It becomes a connected state.

Next, the operation of this embodiment will be described. When the first voice (M = 0) is input, the feature vector calculation unit 1 outputs the feature vector C0 (t) for each frame of the voice.
(Here, t = 1 to T) is obtained and output. The feature vector C0 (t) is stored in the first storage unit 2a of the storage unit 2 and is also input to the delay unit 11 and the average calculation unit 13 of the attached part.

The first storage unit 2a stores the feature vector C0.
Remember (t). On the other hand, the delay unit 11 delays the feature vector C0 (t) for the time required for the average calculation unit 13 to calculate the average value of the feature vector C0 (t). The subtraction unit 12 patterns the CMN-processed cepstrum <C0 (t)> obtained by subtracting the average value of the feature vector C0 (t) from the feature vector C0 (t) input from the delay unit 11. It is output to the matching unit 14. As is well known, the pattern matching unit 14 matches the CMN cepstrum <C0 (t)> with the standard pattern stored in the standard pattern storage unit 6 and outputs the recognition result.

Here, the pattern matching units 5 and 14 use the well-known Viterbi algorithm (Viterbi search), for example, and output from the subtraction units 4 and 12 and the standard pattern stored in the voice pattern storage unit 6. Performs an operation to match with. When the Viterbi algorithm is executed, the optimum state sequence Q for a given observation sequence, that is, the speech input M
You can ask for M.

The recognition result state series storage unit 7 stores, for the voice input M (= 0) obtained by the pattern matching unit 14,
The optimum state sequence Q0 (t) (= {q0 (1), q0 (2), ..., q0 (T)}) is stored.

Next, when there is the next voice input (M = 1), SW1 and SW2 are switched in the direction of the dotted line shown in the figure. The feature vector calculation unit 1 obtains this voice feature vector C1 (t) (here, t = 1 to T) and outputs it in synchronization with the input. This feature vector C1 (t) is stored in the first storage unit 2a.
And is input to the subtraction unit 4. At this time, the feature vector C0 (t) stored in the first storage unit 2a
Has been transferred to the second storage unit 2b, and the normalization parameter calculation unit 3 stores the feature vector C0 (t) stored in the second storage unit 2b and the recognition result state series storage unit 7 in the state vector storage unit 7. The stored optimum state sequence, in other words, the maximum likelihood state sequence Q0 (t)
(= {Q0 (1), q0 (2), ..., q0 (T)}) and using
The equation (3) is calculated to obtain the normalized parameter [CM-1], that is, the cepstrum average value.

[0017]

[Equation 3] Here, ω (qM-1 (t)) is defined as the average value of the average cepstrum of state qM-1 (t) minus the average cepstrum of long-term speech, and is learned and stored as one parameter of the standard pattern. It is what The first term on the right-hand side of the above equation has the content of the utterance + line characteristic of the voice input (M = 0), and the second term has the content of the estimated value of utterance-long speech. Is. For example, voice input (M =
When 0) is a short voice with only "A", the first term on the right side is "A" + line characteristic, and the second term is
"Ah" -Sounds for a long time. Therefore, the cepstrum average value [C0] output from the normalization parameter calculator 3
Becomes like the following formula.

[C0] = {"a" + line characteristic}-{"a" -long time voice} = {line characteristic} + {long time voice} Next, the subtracting section 4 performs CMN according to the following equation. Cepstrum <C1 (t)>. Now, the voice input (M = 1)
As an example, if "Yes", the output from the feature vector calculation unit 1 is "Yes" + line characteristic, so <C1 (t)> is as follows.

<C1 (t)> = C1 (t)-[C0] = {C1 (t) + line characteristic not including line characteristic}-{line characteristic + long voice} = C1 (not including line characteristic) t) -Long duration voice Therefore, according to the present embodiment, it is possible to estimate the cepstrum average value that does not depend on the utterance content and use the cepstrum average value for normalization of the cepstrum average value.

Subsequently, the pattern matching unit 5 causes the above-mentioned <C
1 (t)> is collated with the standard pattern stored in the standard pattern storage unit 6, and the recognition result is output. At the same time, the optimal state sequence Q1 (t) (= {q1 (1), q1 (2), ..., q1 of the recognition result is obtained.
(T)}), which is the recognition result state sequence storage unit 7
Accumulated in. After that, the above-described operation is repeated and the voice recognition is continued.

As described above, according to this embodiment, even if the utterance one sentence before is a short voice such as "A",
The CMN-processed cepstrum <CM (t)> is a long-term voice estimated from the voice at that time-the utterance one sentence before, so that it is not greatly affected by the length of the utterance one sentence before, and the voice recognition rate is prevented from lowering. You will be able to.

Next, a second embodiment of the present invention will be described with reference to the block diagram of FIG. This embodiment is
As in the first embodiment, the voice recognition is performed in real time by using the partial recognition result of the voice that is being uttered halfway, instead of using the utterance of a sentence before.

In FIG. 2, reference numeral 8 is a long-term voice CM0 previously obtained from the learning data, and other symbols are the same as or equivalent to the same symbols in FIG. This long voice CM
As will be apparent from the description below, 0 indicates a short time, for example, one or several frame periods of the voice input M (for example, X1 or X1 to Xt (t <T
s) input period)) is used for the calculation of the normalization parameter, and after this period, it is disabled.

Next, the operation of this embodiment will be described with reference to the flowchart of FIG. In step S1, at the same time when the voice input is started to be received, the feature vector calculation unit 1 causes the feature vector C0 (t) (here,
t = 1 to T) are calculated and sequentially output. These feature vectors C0 (t) are sequentially stored in the storage unit 2a and input to the subtraction unit 4. In step S2, the maximum likelihood state sequence Q
Based on 0 (t), it is determined whether or not several frames of voice have been input. In other words, it is determined whether or not the time Ts corresponding to the several frames has elapsed. When this determination is negative, the process proceeds to step S9, and CM0 is used for the calculation of the normalization parameter. That is, [C0 (t-1)] = CM0
Becomes

In step S4, the normalization parameter calculation unit 3 uses the following equation (4) to sequentially store the feature vector for each frame stored in the storage unit 2a and the pattern matching unit 5 described later. The normalization parameter [CM (t-1)], that is, the cepstrum average value that does not depend on the utterance content is estimated from the input state sequence.

[0026]

[Equation 4] Next, in step S5, the subtraction unit 4 subtracts the normalization parameter [C0 (t-1)] from the feature vector C0 (t),
The CMN-processed cepstrum <C0 (t)> is obtained. In step S6, the pattern matching unit 5 matches the CMN-processed cepstrum <C0 (t)> with the standard pattern stored in the standard pattern storage unit 6, and outputs the recognition result. Also,
At the same time, in step S7, the obtained state series Q0 (t)
Are sequentially sent to the normalization parameter calculator 3. In step S8, it is determined whether or not the voice recognition is completed,
If this judgment is negative, the process returns to step S1 and the calculation of the feature vector is continued. In step S2, it is determined whether the voice input has reached several frames.

When the voice input becomes several frames and the determination in step S2 becomes affirmative, the process proceeds to step S3, in which the normalization parameter calculation unit 3 disconnects the CM08 and calculates the normalization parameter by the state sequence. Use Q0 (t-1). Then, in step S5, the subtraction unit 4 obtains the CMN-processed cepstrum <CM (t)> using the normalization parameter obtained in step S4. Next, in step S6, the CMN-processed cepstrum <CM (t)> is collated with the standard pattern stored in the standard pattern storage unit 6, and the recognition result is output. Also, in step S7,
The state series of the recognition result obtained at the same time in step S6 is output to the normalization parameter calculation unit 3.

After that, the above-mentioned operation is repeated, and the voice recognition for the voice input M is continued.

As described above, according to the present embodiment, the cepstrum average value can be estimated by using the partial recognition result of the voice that is being uttered halfway up to now, without using the utterance of a sentence before. It is possible. That is, it is possible to estimate the cepstrum average value used for normalization at time t using the input up to time (t-1), and real-time speech recognition is possible.

Next, a third embodiment of the present invention will be described.
In this embodiment, in the recognition result of the pattern matching unit 5, only the part determined to have voice is used for estimation, and conversely, the silent part | Q |
Is not used for estimation. According to this embodiment, the risk of misrecognition due to an estimation error can be greatly reduced.

The arithmetic expression of the normalization parameter arithmetic unit in this embodiment is as follows.

[0032]

[Equation 5]

[0033]

As is apparent from the above description, according to the present invention, it is possible to estimate the cepstrum average value with high accuracy even with "short utterance". Even if the utterance is short, the voice recognition rate is not reduced, or the reduction of the recognition rate can be reduced.

Further, since the average value of the cepstrum can be estimated using the utterance up to the middle of the present, it is possible to recognize the voice with a high recognition rate in real time.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a second exemplary embodiment of the present invention.

FIG. 3 is a flowchart illustrating the operation of the second embodiment.

FIG. 4 is a block diagram showing a configuration of a conventional example.

[Explanation of symbols]

1 ... Feature vector calculation unit, 2 ... Storage unit, 3 ... Normalization parameter calculation unit, 4 ... Subtraction unit, 5 ... Pattern matching unit, 6 ...
Standard pattern matching unit, 7 ... State series storage unit of recognition result.

Continuation of the front page (56) Reference JP-A-7-191689 (JP, A) JP-A-8-202385 (JP, A) JP-A-9-90990 (JP, A) JP-A-7-334184 (JP , A) Shingo Kuroiwa, Dieu Tran, Tsuneo Kato, Fumihiro Yato, Study on Real-time Cepstral Mean Normalization Considering Speech Content, Proceedings of the Acoustical Society of Japan, The Acoustical Society of Japan, September 1997, 1997. Autumn I, 159-160 (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 15/02 G10L 15/20 G10L 21/02

Claims (5)

(57) [Claims] 1. An acoustic analysis method for obtaining a feature vector from an input voice, normalizing the feature vector, and performing voice recognition by pattern matching with a standard pattern of voice stored in advance, comprising: an average value of the feature vector; , A part dependent on the utterance content estimated by using the recognition result of the pattern matching is canceled out to estimate a cepstrum average value independent of the utterance content by canceling the part dependent on the utterance content, and the cepstrum average value. Is used for normalizing the feature vector to perform speech recognition. 2. The acoustic analysis method according to claim 1, wherein the estimation of the portion depending on the utterance content is performed by using a recognition result of the pattern matching of the utterance of one sentence before. Analysis method. 3. The acoustic analysis method according to claim 1, wherein the estimation of a portion depending on the utterance content is performed by using a partial recognition result of a voice which is currently uttered halfway. Acoustic analysis method. 4. The acoustic analysis method according to claim 3, wherein the estimation of the cepstrum average value is started after a lapse of a predetermined time from the start of voice input. . 5. The acoustic analysis method according to claim 2, wherein the estimation of a portion depending on the utterance content is performed using only the portion determined to have a voice in the recognition result of the pattern matching. An acoustic analysis method characterized in that