JPH0398098A - Voice recognition device - Google Patents

Abstract

PURPOSE:To enable high-performance phoneme recognition by using a phoneme group result obtained by a segmentation method and combining it with a phoneme discriminating method. CONSTITUTION:The device consists of an amplifier 1, a low-pass filter 2, an A/D converter 3, and a processor 4. The amplifier 1 amplifies an input voice signal and the low-pass filter 2 removes folded-back noises from the voice signal. The D/A converter 3 samples and converts the voice signal into a digital signal and the processor 4 has a computer 5, magnetic disks 6, a terminal, etc., 7, and a printer 8. The computer 5 detects the section of a phoneme and its phoneme group by the segmentation method according to the digital signal of the sampled voice which is inputted from the A/D conversion part 3 and combines the phoneme group with the phoneme discrimination result. Consequently, the high-performance phoneme discrimination becomes possible.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a speech recognition device, and in particular to a speech recognition device that performs segmentation of input speech for each phoneme group, and describes this segmentation method and a phoneme identification neural network applied thereto. This invention relates to a speech recognition device that recognizes phonemes by fusing them.

[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 spotting 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, resulting in highly accurate phoneme recognition. I can't get the rate.

Furthermore, the segmentation method is only used to identify phoneme intervals, and the T8 rhyme groups obtained by the segmentation method are used to determine the final phoneme recognition result, and the proposed method aims to improve the phoneme recognition rate. It has not been. In addition, the latter method has the disadvantage that there are many phoneme recognition errors and intrusion errors near the boundaries of consecutive phonemes, and as a result, a high phoneme recognition rate cannot be obtained.

Therefore, the main purpose of this invention is to solve the misrecognition of phonemes caused by segmentation, the misrecognition of phonemes at phoneme boundaries and insertion errors caused by the phoneme spotting method, and to solve the problems of phoneme recognition and insertion errors at phoneme boundaries using the phoneme spotting method. It is an object of the present invention to provide a speech recognition device capable of determining a phoneme recognition result and achieving high phoneme recognition.

[Means for Solving the Problems] The present invention is a speech recognition device that recognizes input speech, which includes a detection means for detecting a position or section of each preset phoneme group from the input speech, and an input speech recognition device. an identification means for identifying phonemes within a preset phoneme group from the recorded speech using a phoneme identification neural network, and a position or section of each phoneme group detected by the detection means, and an identification means. Speech recognition is performed based on the phonemes identified by.

More preferably, the detection means detects the magnitude of the power in a certain frequency range of the input voice and the power level in a certain frequency range.
1) Segmentation results and their phoneme groups based on acoustic features such as the amount of change in power in the frequency band, the amount of change in the spectrum in a certain frequency band, and the ratio of power in a certain frequency band and another certain frequency band. is determined, and based on the phoneme groups determined as a result of segmentation, the phoneme groups to be applied are narrowed down for the phoneme identification neural network, and by applying the g-rhyme identification neural network according to the narrowed down phoneme groups. Phonological recognition is performed.

Still more preferably, the detection means detects the magnitude of power in a certain frequency band of the input voice, 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. Estimating the segmentation and its phonological group based on acoustic features such as the ratio of power in the band and some other frequency band,
The phonological discrimination neural
A network is applied to perform phoneme identification, and a final phoneme recognition result is determined by using a function representing the validity of the phoneme identification result and the segmentation candidate phoneme group obtained by the detection means.

Still more preferably, the detection means detects the magnitude of power in a certain frequency band of the input voice, 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. A phonological identification neural network that estimates segmentation and its phonological groups based on acoustic features such as the ratio of power between and another certain frequency band, and narrows down the phonological groups using the phonological identification neural network applied by the phonological groups of the estimated segmentation results. Then, phoneme identification is performed by applying a phoneme identification neural network corresponding to the narrowed-down phoneme groups, and the final phonological recognition results are determined.

[Operation] The speech recognition device according to the present invention uses the phoneme group results obtained by the segmentation method and combines them with the phoneme identification method to determine the final phoneme recognition result and perform phoneme recognition. 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. Amplifier 1 amplifies the input audio signal,
The low-pass filter 2 removes aliasing noise from the amplified audio signal.

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

FIGS. 2 to 5 are diagrams showing various methods of recognizing phonemes and recognizing speech according to the present invention.

First, common configurations in each of the methods shown in FIGS. 2 to 5 will be explained. Each of the systems shown in FIGS. 2 to 5 consists of three parts, each consisting of a phoneme segmentation part, a phoneme identification part, and a phoneme determination part. These specific explanations were explained in detail in the patent application (Japanese Patent Application No. 1-61928) previously filed by the inventor of the present invention, and will be briefly explained here. The phoneme segmentation section is performed on a rule basis, and phoneme candidates are detected using global acoustic features on the spectrogram for each phoneme class, and the rough positions of 71 possible phonemes are detected. The phonic classes here include, for example, voiceless fricatives and voiced fricatives.

Next, a hypothesis of the phonological environment is made. That is, for each detected phoneme candidate, the types of sounds before and after each are hypothesized. Next, phonological boundaries are detected and the hypothesis is verified under the hypothesis of the phonological environment. Under the correct hypothesis,
A high degree of confidence is obtained for each hypothesis, and as a result, the phonological environment is detected. Conversely, if the hypothesis is incorrect, the confidence level will be low;
It is not possible to obtain a phonological environment. Judging whether the hypothesis is correct or not depends on the acoustic features on the spectrogram, that is, the magnitude of the power in the lower range, the amount of change in power, the amount of change in the spectrum, and other factors. Judgment is made based on acoustic characteristics such as the ratio of power to the frequency range. Next, the phoneme boundary that gives the maximum confidence for each hypothesized phoneme class is taken as the segmentation result, and the beginning and end of that phoneme and the phoneme class are output with confidence.

FIG. 6 is a diagram showing an example of a time delay neural network (TDNN) for identifying phonemes. next,
With reference to FIG. 6, a method for identifying the phoneme of the segmentation detected as described above will be described.

The time-delay neural network shown in Figure 6 is 1
The 8 consonants are voiced plosives, voiceless plosives, nasals, and voiced fricatives. The sounds are grouped into six classes: voiceless fricatives and flowing sounds, and each group is input to the input layer 11. Input layer 1
1 identifies segmented phonemes by learning the conventionally known pack propagation. Identification of each class is performed by intermediate layer 12.

In this embodiment, the learning of the time-delay neural network is performed by adjusting the end positions of all consonants to positions two-thirds of the way from the end of the input layer 11. Similarly, in phoneme identification, the end of the segmentation result is the input layer 11.
is applied to the same position, and the phoneme that gives the maximum confidence output by the output layer 13 of the time-delayed neural network is taken as the identification result.

The phoneme determination unit shown in Figures 2 to 5 uses the segmentation results for each phoneme class and the phoneme identification results output from the time-delayed neural network applied to that section to identify the phoneme that gives the maximum confidence and its The interval is determined.

The method shown in FIG. 2 identifies phonemes and recognizes speech by combining the simplest segmentation method and phoneme identification method. After the input speech is analyzed and feature extraction is performed, a determination is made in the segmentation unit that, for example, the confidence of voiceless fricatives is 0.62 and the confidence of voiced fricatives is 0.51.

Then, an unvoiced fricative with a high degree of certainty is selected, and this unvoiced fricative is input into the time-delayed neural network shown in FIG. Identification is performed and phonological recognition is performed.

In the example shown in FIG. 3, phonemes are identified and speech is recognized by a means that uses a segmentation method to narrow down phoneme groups. In this example, the input speech is analyzed and as a result of feature extraction, the segmentation part determines the result that gives the maximum confidence, and then identifies voiced consonants depending on whether the phoneme group is a voiced group or an unvoiced group. A time delay network or a time delay neural network for unvoiced consonant discrimination is selectively applied to perform phoneme discrimination within the interval.

In general, the discrimination ability of a time-delayed neural network increases as the number of types of phonemes to be identified increases. Therefore, if there is no confusion between phoneme classes in the segmentation results, a time-delay neural network that performs phoneme identification for each class is used. ) is expected to improve the identification rate. That is, the segmentation unit narrows down phoneme classes and identifies phonemes within that class.

FIG. 7 is a diagram showing an example of the time-delay neural network for identifying voiced consonants and the time-delay neural network for identifying unvoiced consonants described in FIG. 3. The neural network for identifying voiceless consonants shown in FIG.
h) and includes an input layer 21, an intermediate layer 22, and an output layer 23. In addition, the time-delay neural network for identifying voiced consonants shown in FIG. 7(b) is
It identifies consonants (b, d, g, m, n, r, z), and includes an input layer 31, an intermediate layer 32, and an output layer 33.

The example shown in Figure 4 recognizes speech by identifying phonemes using a function representing the validity of the phoneme groups of the segmentation method and the results of the phoneme identification method. Similar to the described embodiment, the segmentation unit determines the reliability of voiceless fricatives and voiced fricatives, and then uses the time-delay neural network shown in FIG. 6 to identify the phonemes within that interval. Phonological recognition is performed. In other words, in the example shown in FIG. 4, phoneme interval candidates and their phoneme groups are output, and the validity of the phonemes identified by the time-delayed neural network and the phoneme classes of the segmentation results can be taken into account. It is expected that both segmentation and phoneme discrimination abilities will improve.

Here, as an example of a function expressing its validity, we use the following equations (1) and (2) to obtain the maximum confidence (
This is a method of using a phoneme that gives a certain factor as a recognition result.

CPrec-combine (CPseg, CF
nn) = (1) CFnn-k aWnn l'
(arg(seg), arg(nn)) (2) However, CPrec: Confidence level of final phoneme recognition CPseg: Confidence level of segmentation result CPn
n: Confidence of phoneme identification result Wnn: Output value of the identified phoneme of the time-delayed neural network arg (seg): Phoneme class of the segmentation result arg (nn): Identification phoneme of the time-delayed neural network k: Coefficient (single-delayed) The reliability of the neural network: The larger k is, the more reliable the output result of the time-delayed neural network is.) ``(): Function indicating the validity of the identified phoneme and phoneme class. If the identified phoneme of the time-delay neural network belongs to the phoneme class of the segmentation result, 1.0 is given, otherwise it is given 1.0, and if voiced/unvoiced sounds match, 0.5 is given.

combine ( ): MYC I N confidence calculation model The example shown in Figure 5 selects the phoneme identification means by using the segmentation method to narrow down the phoneme groups, and then selects the phoneme identification method using the segmentation method to narrow down the phoneme groups. This method identifies phonemes and recognizes speech by using a function that represents the validity of the results of the method.

FIG. 8 is a table showing consonant recognition results by each method of the present invention. When using a time-delay neural network for identifying 18 consonants and two time-delay neural networks for voiced/unvoiced sounds, we consider the validity of the phonology identified by the time-delay neural network and the phonological class of the segmentation results. We conducted an experiment in which conditions were changed, such as how much to trust the output results of the time-delayed neural network when considering the validity of the results. In Figure 8, 18-CO
NS-TDNN shows the case where an 18 consonant discrimination time delay neural network is used, V/UV-TDNN shows the case where two time delay neural networks for voiced/unvoiced sounds are used, and No COMB shows the case where a time delay neural network is used. - Indicates a case where the validity of the network's identified phoneme and the phoneme class of the segmentation result is not considered, and with COMB indicates a case where it is taken into account.

The dependence of Equations (1) and (2) above on the time-delay neural network is k-0.4.
, 0.8 were used. The larger k is, the more reliable the output result of the delay neural network is. Recognition Rate indicates when both phoneme segmentation and phoneme identification are performed correctly, and Insertion Error R
ate indicates the additional doubt rate, Segmentatio
nRate indicates the rate at which the beginning/end boundary error of a phoneme is detected within 50 msec and is judged to have been correctly segmented, and BoundaryAlignment
Error indicates the deviation of the correctly detected boundary from the inspection label, within Correct
The Segmentation Rate indicates the phoneme identification rate at ψ of a segment correctly segmented according to the present invention. The table shown in Figure 8 shows the effectiveness of the method of applying the time-delay neural network after narrowing down the phonological classes, and the validity of the phonological classes identified by the Samurai-delay neural network and the segmentation results. This shows the effectiveness of Hohiro considering the above.

Note that the narrowing down of phoneme groups is not limited to voiced/unvoiced sounds, but can also be divided into fricatives, nasal sounds, plosives, etc., and the speech identification method is applied according to this classification. do it.

Further, in the phoneme identification method of the above-described embodiment, a time-delay neural network is used, but other general statistical methods may be used to identify phonemes within a phoneme group. 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. Identification methods etc. can be applied.

[Effects of the Invention] As described above, according to the present invention, phoneme sections and their phoneme groups are detected from input speech using a segmentation method, and by combining the phoneme groups with the phoneme identification results. To enable high-performance phoneme recognition, and to be able to detect phoneme misrecognition due to segmentation errors and phoneme misrecognition and insertion errors at phoneme boundaries using the phoneme spotting method, and as a result, to enable high-performance phoneme recognition. I can do it.

[Brief explanation of drawings]

FIG. 1 is a schematic block diagram of the entire speech recognition device to which an embodiment of the present invention is applied. FIG. 2 is a diagram showing an example of speech recognition by identifying phonemes by a combination of the simplest segmentation method and phoneme identification method in an embodiment of the present invention. FIG. 3 is a diagram illustrating an example of speech recognition by identifying phonemes using a segmentation method to narrow down phoneme groups. FIG. 4 is a diagram showing an example of speech recognition by identifying phonemes by using phoneme groups of the segmentation method and functions representing the validity of the results of the phoneme identification method. Figure 5 shows that phonemes are identified by means of the segmentation method used to narrow down phoneme groups, and phonemes are identified and speech is identified by using a function that shows the validity of the phoneme groups of the segmentation method and the results of the phoneme identification method. It is a figure which shows an example of recognition. FIG. 6 is a diagram showing an example of the time-delay neural network for identifying 18 consonants used in FIGS. 2 and 4. FIG. 7 shows the time-delay neural network for consonant identification by voiced/unvoiced sounds used in the embodiments of FIGS. 3 and 5.
It is a figure showing an example of a work. FIG. 8 is a table showing the phoneme recognition results according to each method of the present invention. 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, 8 is a printer, 11, 21
．． 31 is an input layer, 12, 22, 32 is an intermediate layer, 13, 2
3.33 indicates the output layer.

Claims (4)

[Claims] (1) 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; an identification means for identifying phonemes within the phoneme group using a phoneme identification neural network, based on the position or section of each phoneme group detected by the detection means and the phoneme identified by the identification means. A voice recognition device characterized in that it performs voice recognition using a voice. (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. The segmentation result and its phonological group are determined based on the acoustic characteristics such as the power ratio between 2. The speech recognition device according to claim 1, wherein phoneme recognition is performed by narrowing down phoneme groups and applying a phoneme identification neural network according to the narrowed down phoneme groups. (3) 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. The segmentation and its phonological group are estimated based on the acoustic characteristics such as the power ratio between A network is applied to perform phoneme identification, and a final phoneme recognition result is determined by using a function representing the validity of the phoneme identification result and the phoneme group of the segmentation candidate obtained by the detection means. 2. The speech recognition device according to claim 1, wherein the speech recognition device performs phoneme recognition. (4) 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. The phonological identification neural network estimates the segmentation and its phonological group based on the acoustic characteristics such as the power ratio between After narrowing down the phoneme, phoneme identification is performed by applying a phoneme identification neural network according to the narrowed-down phoneme group, and a function representing the validity of the phoneme identification result and the phoneme group of the segmentation candidate obtained by the detection means is used. The method is characterized in that the final phonological recognition result is determined by performing phonological recognition.
A speech recognition device according to any one of claims 1 to 3.