JP2008145996A - Speech recognition by template matching using discrete wavelet conversion - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide technique for recognizing a speech by using discrete wavelet conversion for converting data by resolution, and template matching having a keen selecting function. <P>SOLUTION: Phonemes are recognized using template matching wherein a sample cut within a time range of about the pitch of vocal chord vibrations from a peak value and a speech waveform to be recognized are standardized with a maximum amplitude and converted into a coefficient of a discrete wavelet and the coefficient is used as a feature vector. Syllables are recognized by template matching wherein a sample of short syllable level and a speech to be recognized are sampled within the same time range and addition values of absolute values by scales of discrete wavelet coefficients obtained by standardizing waveforms with the maximum amplitude are used as feature vectors. Phonemes of a continuous speech are sectioned by using the fact that the ratio of the addition values by the scales of absolute values of main discrete wavelet coefficients approximate 1 in a phoneme transition area. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a technique for recognizing speech by performing template matching at various resolutions using discrete wavelet transform that converts data according to resolution.

  Currently, keyboard input is used as a human interface for computers and the like that have become widespread due to miniaturization, high performance, and low price. However, keyboard input is less familiar to humans than voice input. Recently, cellular phones have been developed, and voice recognition technology that can be used without being restricted by time and place has been demanded. However, the current mainstream of speech recognition is for mass production that recognizes unspecified speaker's speech using statistical processing, and it has been a challenge to extract universal features from unique speech.

  Personalized voice recognition can be realized by template matching using personal voice as a template. In general, template matching provides sharp selectivity, but the high selectivity impairs the alignment margin. Therefore, if data is arranged by resolution by discrete wavelet transform used for compressing image information such as JPEG2000, template matching can be performed at a resolution suitable for extracting target information.

  Conventionally, it has been considered that it is difficult to cope with voice recognition performed directly by template matching. There have been examples of using wavelet transforms for speech recognition, but they have been tried from a statistical point of view, such as performing wavelet transforms instead of Fourier transforms. There was never. The present invention performs speech recognition processing using speech discrete wavelet transform as preprocessing for template matching.

  The speech recognition algorithm based on template matching according to the present invention is based on a model of a cranial nerve network. In other words, the brain is organized by nerve cells having a template matching function, which is a very sharp selective filter with multiple resolutions. Speech is two-dimensional data whose frequency components change over time, and the data on the pattern appears when the activities of multiple components accompanying speech are transferred as impulses through the neural network. When the impulse pattern matches the pattern when the wiring is connected, the nerve cell generates an impulse again. The impulse is a unit of activity, the sensory cell is activated, the nerve cell is activated and the muscle cell is activated, and the meaning of the unit of activity is borne by the outside world and the activity of each cell itself. In a living body, information is processed by transferring an activity unit called impulse, and voice recognition is also quantized into an impulse activity unit and can be processed digitally.

  In the recognition process, it is a problem to collect information necessary for improving the function of discrimination while being economical and speeding up, and to remove unnecessary information. In the recognition process of the template matching method, a function is determined as to which template having a feature vector is used for matching. When performing speech recognition by template matching, a guideline for setting an algorithm for how to extract what kind of speech features is required, including normalization of pattern data to be collated.

  Actual voices vary widely in utterances themselves. In template matching in which detailed features are performed over a wide range, the number of templates increases and processing becomes difficult. Therefore, as shown in FIG. 1, the feature extraction processing of a speech having a narrow time domain such as a phoneme, the feature extraction processing of a middle time domain speech such as a syllable, and the processing of segmenting a phoneme are performed separately. The data collected by a plurality of methods from the voice of the sample and the data collected by the same method from the newly input voice are collated and recognized by template matching.

  The speech recognition process is performed in accordance with the characteristics of the speech. In other words, phoneme analysis is performed using a wavelet coefficient as a component of a template feature vector for voice feature extraction of about 10 milliseconds in the pitch period of vocal cord vibration. In the feature extraction of the speech waveform in the short syllable utterance period range of about 200 milliseconds as the operation unit of the voicing organ, it is recognized by template matching using the component amount of the wavelet coefficient for each scale corresponding to the change amount for each cycle as a feature vector.

  In order to convert the data into the data arranged in the order of the position according to the resolution, the data arranged in the order of the position according to the scale of the Haar discrete wavelet transform is used. As shown in FIG. 2, if a wavelet coefficient of a high resolution scale is removed stepwise and inverse transformation is performed, data such as a low resolution still image is obtained stepwise. Haar In this wavelet transform, except for the time slot, 0 is set, and the divided waveform is a pair of positive and negative rectangles as the mother wavelet function. The wavelet coefficient is converted and added to the latter half of the data in the time slit. Can be processed in a short time. However, in this wavelet transform, the number of data to be processed is a power of two.

  In the template matching determination process, the degree of coincidence is evaluated by the Hamming distance (sum of absolute values of differences) in which the calculation time is shorter than the Euclidean distance and the difference is remarkable in the distance.

  The speech actually uttered requires more voicing symbol templates than the phonetic symbol types. The short template can be used in common, so it is recognized using the shortest template possible.

  When performing template matching of a speech waveform for the pitch period of the vocal cord vibration for phoneme recognition, the waveform is cut out with the peak value as the starting point and the length as the processing unit within the pitch period, and the wavelet coefficient as the feature vector component Recognize phonemes by template matching.

  When the processing unit is expanded to the entire syllable, a component that expands and contracts with each utterance is included in the syllable, and the number of templates must be greatly increased. The beat (mora), which is the unit of Japanese utterances such as “Iroha ...”, is uttered in about 200 msec, and is a cycle in which the absolute value of the wavelet coefficient of the entire short syllable waveform uttered quickly is added for each scale. Short syllable recognition is performed by template matching using a value corresponding to the component amount for each band as a feature vector component.

  As preprocessing for feature extraction for syllable recognition, utterance is divided and collected, samples of short syllable units are collected, the amplitude is normalized to 1 for the maximum amplitude of the collected section, and data to be collated from the input speech is also included A feature vector is obtained by sampling with the same frame length as the sample and normalizing in the same manner.

  It is difficult to recognize phonemes in regions where phonemes transition in continuous speech. Therefore, the component amount for each scale of the discrete wavelet transform is obtained and the phoneme transition is detected with the ratio. [FIG. 3] shows a waveform of a voice in which “I” is uttered briefly but “Aiueo” is uttered continuously. FIG. 4 shows the ratio of the main component amounts by scale of the discrete wavelet transform for this sound. In FIG. 4, the ratio of the main component amounts by scale is close to 1 in the region where the phoneme transitions. Using this method, syllable segmentation of continuous speech can be studied.

  If a phonological symbol string as information processing of speech recognized from speech is converted into a registration number and an organization for decompressing it is constructed, speech information can be compressed and transmitted or stored. Since the digital signal is narrowed down to one by the AND circuit and can be developed into a plurality of components by the OR circuit, the collected digital data pattern can be converted by the digital circuit. In that case, the inventor's Japanese Patent Application No. 2004-217828 writable bidirectional logic circuit is convenient. In this case, output coding is written as a decoder from the reverse direction, and the conduction point group of the circuit is used as an encoder.

  The current high-performance personal computer workstation has a register structure for processing 64 bits, so that a technical environment for processing a speech language is in place. Incidentally, 256 phonetic symbols can be distinguished by an 8-bit digital signal. Therefore, if a word is expressed by 8 phonetic symbols, it is specified by 8 × 8 = 64 bits of word phonetic symbol data. When the input is performed in units of 64-bit words and 8,192 words are written, the matrix points are 524,288. If 8,192 words are specified by registration numbers, they can be specified by 13 bits. Here, the registration number of the output can be determined by a registration number counter. If the output is expressed by nine 128-bit, 7-bit characters, the character output is 63 bits.

  In the information processing apparatus, if the pattern matching is hierarchized without converting the voice information into the character information, the information can be compressed and processed. That is, the phoneme identification information is converted into a word level registration number, and the word level registration number string is converted into a sentence level registration number. Conversely, the sentence level registration number is returned to the original word level registration number string, and the word level registration number is converted into a phoneme symbol string. If one word is specified by 13-bit data, a combination of nine words 13 × 9 = 117 bits is input as one sentence, and if 8,192 kinds of sentences are registered, the input side The point of the matrix is 958.464. Since the sentences classified into 8,192 types can be distinguished by 13-bit numbers, the matrix elements on the output side are 106,496. A circuit of this scale can be realized with a semiconductor integrated circuit.

  In the present invention, a notebook computer with an AD conversion card inserted is used to examine the data before and after processing in Microsoft Office Excel, and the software is programmed with Visual Basic for Application (VBA). It can be used as software embedded in the input part of a personal computer or robot.

  In order to process a large amount of template matching at high speed, if templates are written in hardware, they can be matched in parallel. It is best to produce the organization of the present invention with programmable semiconductor integrated circuits. To construct a semiconductor integrated circuit, the presence of an active unit is transferred as the presence or absence of charge, and the charge is transferred by a CCD or dynamic MOS IC circuit. As a circuit that realizes a data conversion function by transferring an activity unit, there are an inventor's patent No. 3496065 impulse electronic device and an inventor's Japanese Patent Application No. 2004-217828's writable bidirectional logic circuit.

  The implementation method of the present invention will be described below through an embodiment of speech recognition of a specific speaker by template matching using discrete wavelet transform.

  Similarly, since the uttered voices coincide with each other, the sample may have a similar waveform segment. FIG. 5 shows the Hamming distance of template matching with five types of waveforms cut out 12.8 msec from the voice waveform itself of FIG. 3 continuously uttered “Aiueo”. [FIG. 5] shows a state in which the continuous voice continuously changes the degree of coincidence with the template.

  A shorter waveform segment requires less processing unit data and a shorter processing time. When the number of templates is small, the processing time can be shortened when processing with a computer. For this purpose, the sample template section used for collation is made short and has a low resolution. FIG. 6 shows the Hamming distance of template matching with 15 feature vectors obtained by cutting the waveform of speech continuously uttered “Aiueo” from its own speech by 6.4 milliseconds. [FIG. 6] shows that it is possible to identify even with a collation with low selectivity.

  Since there are many templates, processing time is required, so we want to collect vowel sounds that can be used as common samples. The pitch of the vocal cord vibration differs depending on the type of vowel and the way of pronunciation, and the waveform changes as the pitch changes. [Fig. 7] shows the change in pitch of the voice when uttering "Uh, uh, oo".

  As an example in which the processing time of the recognition processing is greatly shortened, a discrete wavelet transform is used in which a waveform sample obtained by cutting 6.4 milliseconds from a peak value from a voice waveform having a pitch of 9.5 milliseconds is used as a common sample. An example of recognition of phonemes of a specific speaker by template matching will be shown, and a guideline for producing this recognition method will be described.

FIG. 8 shows the template matching distance between the voice “Aiueo” and the voice uttered continuously and the vowel sample “Ah, ii, woo, uh, oo”. Here, five kinds of vowel samples were cut out by 6.4 milliseconds from a waveform having a pitch of 9.5 milliseconds in a voice waveform uttered. Note that the input speech “I” is uttered briefly, and the sample “I” uttered “Ii” cannot be recognized.

  FIG. 9 shows the template matching distance between the voice uttered “Kakikukeko” and cut from the peak value by 6.4 milliseconds and the same vowel sample as the process of FIG. The vowels of “ko” are recognized as samples of “u”. It is thought that it will be recognized if the voice of “ko” is rounded up quickly. Multiple templates are required for the same phonetic symbol. Note that the K sample waveform was cut off by 6.4 milliseconds from the leading portion characteristic of the consonant of the “ki” waveform. Since the waveform of the head region as the feature of K does not have vocal cord vibration and the consonant is related to the whole utterance operation, the feature should be extracted and recognized from the speech segment expanded as a short syllable.

  FIG. 10 shows the template matching distance between the voice segment cut out for 6.4 milliseconds with the voice uttered “Sashisuseso” and the same vowel sample as in FIG. The vowels of “shi” are recognized as a number of vowels. The utterance of “shi” is considered to be “shi” instead of “si”. Note that the sample of S is a waveform obtained by cutting off a characteristic leading portion as a consonant of the waveform of “shi” by 6.4 milliseconds. The waveform of the head region, which is a feature of S, does not have vocal cord vibrations as in the case of K, and the consonant is related to the entire utterance operation.

  In template matching in which the component amount by scale is a feature vector, it corresponds to the change component by cycle, so that an amount with a small amplitude level has little influence. So, say "Ah, S, T, N," and collate template matching using a scale-based component quantity as a feature vector for the syllable waveform sample in the area covering all syllables [409.6 milliseconds]. As shown in [FIG. 11], the effect appears only by shifting the cut-out position of the waveform from the same voice by 0.8 milliseconds.

  FIG. 12 shows a state in which template matching with a normally uttered syllable is collated by using a short syllable unit of 204.8 milliseconds as a sample and using a component amount by scale in the range as a feature vector. The central part recognizes the syllable and the terminal part recognizes the vowel. When template matching is performed with a voice sample of a short syllable period, it is not necessary to shift the frame as frequently as shown in FIG.

  Although the gist of the present invention has been described above, the present invention is not limited to the conditions of the embodiments shown in the drawings, and modifications and additions within the scope not departing from the gist of the present invention are included in the present invention. . Since there are various voices and various voice recognition uses, when the present invention is actually used, the matters described in this specification are specifically constructed.

  Examples of the use of the organization for inputting speech according to the present invention include speech information compression, speech input for a purchase order such as a restaurant, speech typewriter, speech correction device, automatic translation telephone, industrial robot, and nursing robot.

It is a figure which shows the structure of the audio | voice recognition structure | tissue judged by comparing with the data of the sample about the data extract | collected from the audio | voice input combining combining discrete wavelet transform and template matching. It is a figure which shows the multi-resolution of a discrete wavelet transform by the waveform calculated | required by inverse transformation after removing the wavelet coefficient of a high resolution scale in steps. It is a figure which shows the waveform of the audio | voice which uttered continuously "I" used short in the Example of the analysis of speech using a discrete wavelet transform, and the recognition process as "Aiueo". It is a figure which shows that a phoneme transition can be detected with the ratio of the component amount according to scale of the discrete wavelet transform of the audio | voice shown in FIG. It is a figure which shows the Hamming distance of a template matching with five types of sample waveforms which cut out 12.8 msec from the audio | voice waveform shown in FIG. It is a figure which shows the distance of template matching with five types of low-resolution waveforms which clipped the waveform of the audio | voice shown in FIG. 3 from own audio | voice 6.4 msec. It is a figure which shows the change of the pitch of an audio | voice when uttered "woo, uh, oo" referred when adopting 9.5 msec as a common pitch. FIG. 7 is a diagram showing the template matching distances of the voice shown in FIG. 3 with five types of low-resolution waveforms obtained by cutting 6.4 msec from the waveform of the pitch of 9.5 msec shown in FIG. It is a figure which shows the distance of the template matching of the audio | voice which uttered "Kakikukeko" by five types of low resolution waveforms cut out from the waveform of the pitch of 9.5 msec 6.4 msec. It is a figure which shows the distance of the template matching of the audio | voice which uttered "Sashisuseso" by five types of low resolution waveforms cut out from the waveform of a pitch of 9.5 msec by 6.4 msec. It is a figure which shows the influence which shifted time by 0.8 msec in the distance of the template matching which used the component amount according to the scale of the same syllable [409.6 msec] as the feature vector. It is a figure which shows that a syllable can be recognized by the Hamming distance of the template matching which shifted the frame of 204.8 msec of short syllable units every 6.4 msec, and made the component amount according to the scale of a short syllable unit into the feature vector.

Claims (5)

  Extract the speech in the time range of the pitch of the vocal cord vibration from the peak value, normalize the waveform in that region with the maximum amplitude, convert it to a discrete wavelet coefficient, and use the wavelet coefficient as a feature vector to match the sample by template matching Recognize phoneme level speech by finding a distance measure.   A speech waveform segment is collected in a time range of about a short syllable unit, the waveform in that region is normalized to the maximum amplitude, converted to discrete wavelet coefficients, and the sum of the absolute values of the discrete wavelet coefficients for each scale is a feature vector. Recognize short syllable level speech by obtaining a distance measure to the sample by template matching.   Segmentation of phoneme transitions in continuous speech using the fact that the ratio between the scales of the absolute values of the discrete wavelet coefficients of the main scales constituting the speech is close to 1 in the phoneme transition region.   A method including a plurality of claims 1, 2, and 3, data representing voice characteristics are sampled in a multiple manner, and a template matching distance measure with the sample data is obtained for the multiple samples. Recognize voice.   A software system or electronic circuit device having a function of transmitting, storing, or operating sound characterized by any of the items of claim 1, claim 2, claim 3, and claim 4.

Images