JPH02239290A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To improve the phoneme recognition rate by recognizing a voice based on the position or the section of each phoneme group detected by a detecting means and the phoneme discriminated by a discriminating means. CONSTITUTION:This device consists of an amplifier 1, a low pass filter 2, an A/D converter 3, and a processor 4. The processor 4 consists of a computer 5, a magnetic disk 6, a terminal 7 or the like, and a printer 8, and the computer 5 recognizes the voice based on the digital signal of the voice inputted from the A/D converter 3. That is, the position or the section of each phoneme group is detected by the detecting means, and a phoneme in the phoneme group preliminarily set from the inputted voice is discriminated by the discriminating means. Thus, the phoneme is recognized with a high performance, and a voice recognizing device of high performance is constituted.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a speech recognition device, and in particular, detects the position or interval of each phoneme group of input speech, performs segmentation, and performs segmentation on the detected position or interval. The present invention relates to a speech recognition device that recognizes phonemes.

[Prior art and problems to be solved by the invention] Conventional speech recognition methods include a method in which a continuous speech waveform is segmented by time division, and then phoneme recognition is performed; A so-called phoneme swapping method has been proposed that simultaneously performs segmentation and phoneme recognition of the segment.

However, in the former method, segmentation is performed using a combination of simple parameters such as uniform power and spectral changes regardless of the phonological environment in which each phoneme exists, so it is difficult to perform highly accurate segmentation. I can't. As a result, a high phoneme recognition rate cannot be obtained. Also, in the latter method,
This method has the disadvantage that there are many phoneme recognition errors and insertion errors near the boundaries of consecutive phonemes, and as a result, a high phoneme recognition rate cannot be obtained.

Therefore, the main object of the present invention is to provide a speech recognition device capable of high level phoneme recognition by solving phoneme misrecognition caused by segmentation errors and phoneme recognition errors and insertion errors at phoneme boundaries caused by the phoneme swapping method. It is.

[Means for Solving the Problems] This invention is a speech recognition device that recognizes input speech, comprising a detection means for detecting a position or section of each preset phoneme group from the input speech; and identification means for identifying a tone hammer within a preset phoneme group from input speech.

More preferably, the detection means detects the magnitude of the power of the input voice in a certain frequency band, the amount of change in power in that frequency band, the amount of change in spectrum in that frequency band, and the difference between one frequency band and another frequency band. and a means for detecting a position or section for each phoneme group based on the power ratio between the two phoneme groups.

More preferably, the identification means identify using statistical methods designed to identify phonemes within predefined phoneme groups.

Furthermore, after detecting the position or section of each preset phoneme group, the phonemes within the preset phoneme group are identified.

[Operation] In the speech recognition device according to the present invention, the detection means detects the position or section of each phoneme group, and at the same time, the identification means identifies phonemes within a preset tone group from the input speech. As a result, high-performance phoneme recognition becomes possible, and a high-performance speech recognition device can be constructed.

[Embodiment of the Invention] FIG. 1 is a schematic block diagram of a speech recognition device to which the present invention is applied. Referring to FIG. 1, the speech recognition device includes an amplifier 1, a low-pass filter 2, an A/D converter 3, and a processing device 4. Anbu 1 amplifies the input audio signal,
The low bass filter 2 removes aliasing noise from the amplified audio signal.

The A/D converter 3 converts the audio signal into a 16-bit digital signal using a 12 kHz sampling signal. The processing device 4 includes a computer 5, a magnetic disk 6, and terminals 7.
and a printer 8. The computer 5 performs voice recognition based on the voice digital signal input from the A/D converter 3 using the method shown in FIGS. 2 to 5, which will be described later.

FIG. 2 is a diagram showing a procedure for detecting a section for each phoneme group according to an embodiment of the present invention, FIG. 3 is a diagram showing an example of subekto mouth duram, and FIG. 4 is a diagram showing recognition results. FIG. 5 is a diagram showing an example of identifying phonemes using a neural network.

Next, with reference to FIGS. 1 to 5, a specific operation of an embodiment of the present invention will be described.

A digitized phonetic spectrum is provided to a computer 5 from the A/D converter 3 shown in FIG. The computer 5 performs step (abbreviated as SP1 in the figure) SP
In I, rough phoneme features on the spectrum mouth durum are referred to based on the input phoneme spectrum. Third
The figure is a vectorogram when pronouncing [sukunakutomoJ, where the vertical axis shows the frequency and the horizontal axis shows the passage of time. In this spectrogram, black areas indicate large power, and white areas indicate small power.

In step SP2 of FIG. 2, phoneme candidates are detected. That is, based on the reference result of the phoneme features in step SPi described above, phoneme candidates are detected using the rough features for the rough position of each phoneme group. The phonological groups here include, for example, voiceless fricatives, voiced plosives,
These include nasal sounds and flowing sounds.

In the spectrum shown in FIG. 3, in the interval from 335 msec to 492 msec, corresponding to /S/,
The power in the frequency band from 4000Hz to 6000Hz is large, and the power is small in the frequency band around 1000Hz to 2000Hz, and the cutoff point is 500QH
Since it is near z, it is determined that it is almost a voiceless fricative or a voiced fricative, and the voiceless fricative and the voiced fricative are considered as phoneme candidates.

Next, following /S/, corresponding to /k/, 492
Based on changes in power, changes in spectrum, etc. in the interval of ~562 msec, unvoiced plosives are selected as phoneme candidates.

Next, in step SP3, a hypothesis of the phonetic environment is made. That is, for each phoneme candidate detected in step sP2 described above, types of phonemes and phoneme transformations before and after the phoneme are hypothesized.

That is, the types of phonemes before and after each of the voiceless fricative and voiced fricative detected in step SP2 are hypothesized. For the voiceless fricative /s/, a silence, a stop, and a vowel are hypothesized before it, and a stop, a stop, a vowel, and a fricative are hypothesized for the phoneme after it. Step SP
For the voiced fricative detected in step 2, the types of phonemes before and after are tentatively assumed, with the preceding phoneme being silent and a vowel, and the subsequent phoneme being a vowel.

Possible phoneme boundaries are detected and hypotheses are verified for each phoneme environment hypothesized in step SP3 described above. Under correct hypotheses, a high degree of confidence is obtained for each hypothesis, and as a result, the phonological environment is detected. On the other hand, if the hypothesis is incorrect, the confidence level will be low and the phonological environment will not be obtained. Judging whether the hypothesis is correct or not depends on the acoustic features on the spectrogram, that is, the magnitude of the power in a certain frequency band of the input voice, the amount of change in power, the amount of change in the spectrum, and the power in other frequency bands. It is determined based on the ratio of

In step SP5, among the sections in which each phoneme group has been determined, the section with the highest confidence is determined as the final segmentation and phoneme group result. Based on the final segmentation result, a corresponding phoneme group is identified in step SP6. Step SP
Regarding the hypothesis of silence in 3, the voiceless fricative starts from 335 msec, and its confidence (c f) is 0.6.
4, no result was obtained for the vowel hypothesis, and 325 msec for the stop consonant hypothesis.
Starting from , the result is that the confidence level is 0.60. Furthermore, for the plosive hypothesis, it is hypothesized that the starting boundary of 492 msec is the end of /s/, and the confidence level thereof is 0.66.

In step SP6, the result with the highest confidence is selected, and in step SP7, /S/ is 335 msec.
The end is identified at 492 msec, and segmentation is determined based on this, and at the same time phoneme group identification is performed.

Next, a method for identifying the phoneme of the detected segmentation will be described with reference to FIG. The time-delayed neural network shown in Figure 5 divides the 18 consonants into voiced plosives, voiceless plosives, nasals, voiced fricatives, and voiceless fricatives. The flowing sounds are grouped into six classes, and each group is used as the input layer 10. The input layer 10 is trained by the conventionally known pack blobagation.
Identify segmented phonemes.

Identification of each class is performed by the input layer 11.

The learning of the time-delayed neural network starts from 150 msec before the input layer 10 of the final position of all consonants.
Similarly, in phoneme identification, the end of the segmentation result is applied to the same position of the input layer 10, and the output layer 12 of the time-delayed neural network outputs the phoneme that gives the maximum confidence as the identification result. shall be. FIG. 4 shows an example of this identification result.

Note that in the position detection in the above-described embodiment, phoneme groups and their sections are shown. However, in addition to this method, it is also possible to identify phonological groups with certain characteristics and their characteristics, such as phonological groups with rupture characteristics and rupture positions, phonological groups with local power dip characteristics and dip positions, etc. It is also possible to use a method based on the position of .

Furthermore, in the phoneme identification method shown in FIG. 5 above,
Although a time-delay neural network was used, it is also possible to identify phonemes within phoneme groups using other general statistical methods. For example, phoneme identification methods using general neural networks, phoneme identification methods using HMM, phoneme identification methods using Bayes' rule, phoneme identification methods using linear discrimination, and phoneme identification methods using standard patterns designed using methods such as LVQ. methods etc. are applicable.

[Effects of the Invention] As described above, according to the present invention, the position or section of each preset phoneme group is detected from input speech, and the phonemes within the preset phoneme group are identified. As a result, it is possible to solve the conventional misrecognition of phonemes due to segmentation errors, misrecognition of phonemes at phoneme boundaries due to phoneme swapping, and insertion errors, and improve the performance of phoneme recognition.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the overall configuration of a speech recognition device to which an embodiment of the present invention is applied. FIG. 2 is a diagram showing a procedure for detecting intervals for each phoneme group according to an embodiment of the present invention. FIG. 3 is a diagram showing an example of a spectrogram. FIG. 4 is a diagram showing the voice recognition results. FIG. 5 is a diagram showing an example of voice recognition using a time-delay neural network. In the figure, 1 is an amplifier, 2 is a low-pass filter, 3 is an A/D converter, 4 is a processing device, 5 is a computer, 6 is a magnetic disk, 7 is a terminal, and 8 is a printer. Patent applicant: A.T.R. Co., Ltd. Subject of automatic correction: Figure 4 of the drawing dated August 2, 1999

Claims (4)

[Claims] (1) A speech recognition device that recognizes input speech, comprising: a detection means for detecting a position or pattern for each preset phoneme group from the input speech; and a detection means for detecting a preset phoneme group from the input speech. an identification means for identifying phonemes within the group, and a position or section of each phoneme group detected by the detection means;
A speech recognition device, characterized in that speech recognition is performed based on the phoneme identified by the identification means. (2) The detection means detects the magnitude of the power of the input voice in a certain frequency band, the amount of change in power in a certain frequency band, the amount of change in spectrum in a certain frequency band, and the amount of change in the power in a certain frequency band. 2. The speech recognition device according to claim 1, wherein the position or section of each phoneme group is detected based on the ratio of power in a certain frequency band and a certain other frequency band. (3) The speech recognition device according to claim 1, wherein the identification means uses a statistical method designed to identify phonemes within a preset phoneme group. (4) After the detecting means detects the position or section of each preset phoneme group, the identifying means identifies the phoneme within the preset phoneme group. The speech recognition device according to item 1.