JPS58121100A - Word voice recognition system - 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] <Background of the Invention> The present invention is a word speech sIm method that can simultaneously recognize various word sounds input from a plurality of input devices. Since the semi-patterns were stored in the form of a time series of spectral parameters, the storage capacity required to store them was enormous as the number of single @ became large. When one acupuncture circuit is provided, there is a drawback that the number of words that can be iIgwlc by each recognition circuit is kept small.

<Overview of the Invention> In order to solve these drawbacks, this invention reduces the storage capacity by storing each half-time standard pattern by a combination of an onomatopoeic iI bare quasi-pattern and a word dictionary. By storing a set of words as a standard pattern and specifying which word set is to be recognized when using the device, it is possible to equivalently recognize a large number of words with one device. It is something.

<Embodiment> FIG. 1 shows an embodiment of the present invention, in which input is made to a plurality of input circuits 2. +ia! Internal organ selection circuit 2 has input terminal 1!
A smaller number of recognition circuits 141 to 14m than 1n are connected. The %-sensing circuits 141 to 14m include a spectral parameter extraction section 3, an onomatopoeic standard pattern accumulation section 4, a spectral distance calculation section 5, a word dictionary accumulation section 6, an address specification section 7, and a time normalized spectrum matching section 8. and the word determination section 9. These mm circuits are 14t to 14m each! lli! through the i1gvjt single #tiMll connection output interface unit 10! It is output to the 1111M power output terminal 11. Approved steel department selection circuit 2, recognition circuit 1
41 to 14m1 output interface 7 eighth section 10 is control section 1
It is controlled by 2. The control section 12 is supplied with multiple control signals through a control signal input terminal 13 .

In this single-onomatopoeia recognition method, a spectral parameter set for each onomatopoeic sound is stored in each pseudo-fM standard pattern storage section 4 in advance before use.

These spectral parameters include correlation coefficients, cepstrums, band-pass filter output powers, etc. For a method of creating a quasi-tWt collar pattern, see, for example, Japanese Patent Application No. 139094/1985 invented by Tube, Furui, and Hakode. The method described on the jar can be used. In this method, from a large number of sets of spectral parameters extracted from the speech of one or more speakers, a clustering method is used to generate a set of spectral parameters ranging from several tens to 200 f! Class II patterns are selected and stored in the *f rhyme quasi-pattern storage section 4. In the storage unit 6, each recognition target rate is stored as a continuous series of symbols indicating the fold f-standard pattern.

For this method, for example, the method described in the invention of the above-mentioned tube material, Furui, and Hakode, can be used.

In general, even if there are a large number of igw & target words, at each point in time when a gut action should be performed, the wit of the words that should be candidates for the input speech is often a part of the total recognition vs. breaking single candy. . For example, in a certain service, even if month, day, -day, and place name Cape are to be recognized, at the time when the speaker should say month and day, it is necessary to include up to -day among the candidates and perform recognition production. There is not necessarily a lid. Therefore, the rate of this invention! In the speech recognition method, when storing word dictionaries, they are divided into groups of words to be recognized at each point in time, and for each group, the number of words in the word dictionary is determined. A table indicating whether it is a series or not is created and stored in the p1 address designation section 7.

<M fiber movement> After accumulating the onomatopoeic standard patterns, word dictionaries, and tables showing leather candy groups in this way, the M fiber movement of unknown word sounds is
Move to Makoto. For example, to explain a service system that uses voice M-Makoto to reserve a seat on an airline ticket, first, when an incoming call from a user is detected, the system uses a synthesized voice to reserve a seat. After telling the user, questions are asked in a predetermined order using synthesized voices, and part of the user's response to the question is recorded. At this time, for example, when asking the question "Where are you from?", you only need to recognize the word group of place names (word candidates), and when asking "What day is it?" It is good to only recognize groups.
As mentioned above.

Immediately after asking a question using a synthesized voice, the system inputs a recognition operation start signal to the control signal input terminal 13, and the control unit 12 sets the recognition path 141 to 14rn to a blank state. The selection circuit 2
As a result, the voice input terminals 11 to ln are connected to the idle recognition circuit through the authentication section selection circuit 2. Next, to the control signal input terminal 13,
ilI! ! Enter the name of the single group to be targeted (group code), specify the candidate word set to be targeted, and add the part to the address finger of the ``wgg (ro) path selected as described above. Enter this yearning in 7 and enter this group name (
If you use the address range stored in the address specifying section 7, specify the address range of the series to be recognized in the word dictionary. Input the waveform of a single AT voice. The acoustic waveform input to the audio input terminal may be one that has passed through a telephone line or may be one that has been taken from a microphone. The voice waveform input from the voice input terminal is selected as described above and sent to the spectrum extraction input parameter extraction section 3 in the nine-tone circuit, where it is subjected to spectrum analysis at short time intervals of, for example, lQmg. As for the result, the spectral distance calculation unit 5 calculates the spectral distance from each onomatopoeic pattern that successively follows from the pseudo g@l/A quasi-pattern storage &s4 at short intervals. Using this total xFi rice and the onomatopoeia-pattern series of the specified candidate word set within the specified address range stored in the candy dictionary storage unit 6, a spectral pine tree that absorbs the time expansion and contraction of speech is used. The matching is performed by a time normalized spectrum matching unit 8, and the degree of similarity between the input speech and each sequence is input to a word determining unit 9. The word determination section 9 selects the series with the highest degree of similarity, and outputs the word name (word number) as -oritsumugi-kama to the futsu-yuibutsu output terminal 11 via the silk-tie button output intern ace section 1O. Output.

l1lI for unspecified speakers! As in the case of fibers,
If multiple representative sequences are stored for each recognition target word, select multiple sequences with a high degree of similarity and make a decision based on majority vote regarding the word name.
Highly axial single-strand determination can be performed.

With such a configuration, according to the method of the present invention, each word sound is expressed as a symbol sequence whose units are pseudo f-pattern names (symbols), so the memory St stores spectral parameters. Compared to the conventional method of storing data, it requires significantly less data, and therefore it is possible to store standard forms of many words in each of the plurality of valley & Iim circuits, and to recognize many words. From among this large number of recognition target words, it is possible to arbitrarily set a set of words to be recognized at each time point, so multiple input devices (input circuits)
Inputting the input voice into any recognition circuit that is idle at that time, and performing the 9M-sensing motion with the set word set as the target becomes a kuji-noh. Therefore, compared to the conventional method in which the types of rate times that can be recognized by each recognition circuit are fixed at K, the method of the present invention allows one mR device to be extremely effective against audio from multiple input devices. It can be used theatrically.

<Effects of the Invention> As explained above, according to the single-chain speech one selection method according to the present invention, a set of candidate words to be targeted by each of a plurality of recognition circuits is instructed at each time point, and a set of candidate words to be targeted by each of a plurality of recognition circuits is specified at each time point. The voice of Il! is detected by the recognition circuit in the overloaded state. Since it is possible to use one iron, it is possible to target a large number of words when recording the voice input by many users through telephones or microphones, and the processing efficiency can be greatly increased. It is possible to reduce the probability that a person will be placed in a waiting state. Another word 1M created by this invention! According to the acupuncture method, the distance (similarity) calculation required for matching is only a calculation between the input speech and the onomatopoeic standard pattern, so a spectral parameter series is stored in a standard format for each word as in the past. This method has the advantage of significantly reducing the amount of calculation compared to the conventional method.

[Brief explanation of the drawing]

The figure is a block diagram showing the basic configuration of the word speech recognition method according to the present invention. 11 to 1n: audio input terminal, 2: drinking section selection (b) path,
3: Varnish vector parameter extraction unit, 4: Pseudo-fIII#
1 collar semi-pattern storage section, 5 varnish vector distance calculation section, 6
: word dictionary storage section, 7: address specification section, 8: time normalization spectrum matching section, 9: word judgment section, xo: Guanori connection output interface 7-8 section, 11: gchen result output terminal, 12: Control unit, 13: Control signal input terminal, 141~
14m: Recognition circuit. Patent Applicant Nippon Telegraph and Telephone Public Corporation Agent Kuruma Kusano

Claims (1)

[Claims] (1) In a line-based device that recognizes voice input from multiple input devices, multiple recognition circuits and
The Ig circuit is provided with a dictionary containing a plurality of candidate word sets represented by onomatopoeic standard patterns and symbol strings indicating the same, and speech input from a plurality of input devices is connected to the recognition circuit. Supply it to someone! selects a recognition circuit in an idle state according to input detection to the device, specifies a set of candidate words to be recognized in the recognition circuit, and selects a set of candidate words to be recognized from the set of candidate words in the input speech. A word voice wtllI method that selects the word with the most similar standard pattern and outputs it as a recognition result.