JPS58159592A - Polysyllabic word 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 (a) Technical Field of the Invention The present invention relates to a speech recognition system, and particularly to a multi-syllable word recognition method that does not require mode switching to recognize multi-syllable words and monosyllabic words. .

(b) Technical background The speech recognition system mainly recognizes monosyllabic words, but if multi-syllabic words can also be recognized, special symbols such as line breaks, single periods, commas, r, J, etc., and saves, etc. It becomes possible to input control terms such as "print" simultaneously, and keys for switching modes such as word mode and kana mode are no longer required, which simplifies the configuration of the voice recognition system and improves operability.

Therefore, it is desired to develop a speech recognition system that can recognize monosyllabic words and has a high recognition processing speed for multisyllabic words.

(c) Object of the Invention The object of the present invention is to provide a speech recognition system that does not require mode switching, can recognize multi-syllable words and mono-syllabic words, and has a fast recognition processing speed for multi-syllable words. It is about providing.

(d) Configuration of the Invention The configuration of the present invention includes providing and registering means for non-uniformly sampling voice feature variations to obtain a reduced parameter time series which is a feature pattern of a fixed length, and registering the reduced parameter time series using the reduced parameter time series. The feature parameters of the unknown input speech extracted by the same means are compared to determine whether it is a multi-syllable word or a monosyllabic word. For words determined to be multi-syllable words, the word with the best characteristic parameter is selected as the first candidate, and multiple candidates are sequentially selected. Next, the candidate selection unit narrows down the matching candidates for the plural ff clause words and performs DP matching.

Furthermore, words determined to be monosyllabic words are directly subjected to DP matching by the monosyllabic word recognition unit.

Regarding the means for obtaining the reduced parameter time series, there is a method that efficiently determines recognition target candidates regardless of the content and size of the recognition target word under a relatively simple algorithm (Japanese Patent Application No. 1983).
5-062059) and is described therein. To briefly explain, the audio signal of an unknown input word is analyzed and refined) ji (The input feature parameter time series extracted from No. 8 is divided into at most 10 sections or less, and the feature parameter values in each section are calculated. It is designed to average.

(e) Examples of the invention The figure is a block diagram of a circuit showing one embodiment of the present invention.

First, in order to register speech in advance, the speaker connects the switching unit 3 to the reduced parameter time series extraction unit 4, the multi-syllable word recognition unit 13, and the monosyllabic word recognition unit 14 under the control of the control unit 9, and identifies it as a monosyllabic word. Add multiple a clause words from the input. The preprocessing section 1 performs audio level adjustment, analog-to-digital conversion, etc., and sends the results to the parameter extraction section 2. Parameter extraction part 2
extracts the parameters and sends them to the switching section 3. The feature parameters input from the switching unit 3 to the reduced parameter time series extraction unit 4 are non-uniformly sampled, and a fixed length reduced parameter time series is extracted and sent to the switching unit 5. #Ai for monosyllabic words entered in switching section 5! The small parameter time series is stored in the monosyllable reduced parameter time series storage section 6, and the reduced parameter time series for multi-syllable words is stored in the multi-syllable reduced parameter time series storage section 7. At the same time, the characteristic parameters of monosyllabic words are stored in the monosyllabic word recognition section 14, and the characteristic parameters of multi-syllable words are stored in the multi-syllable word recognition section 13, respectively.

Next, in order to perform speech recognition, the speaker connects the switching section 3 to the reduced parameter time series extraction section 4 and the storage section IO, and connects the switching section 5 to the collation determining section 8 under the control of the control section 9. The input audio is processed by the preprocessing section 1, the parameter extraction section 2, and the like described above. Switching unit 3° reduction parameter time series extraction unit 4.
The data passes through the switching unit 5 and enters the matching determination unit 8 where it is compared with the reduced parameter time series from the single syllable reduced parameter time series storage unit 6 and the multi-syllable reduced parameter time series storage unit 7. When the collation determining unit 8 determines that the word is a multi-syllable word, it selects the word with the most similar characteristic parameters as the -Mth sleeve, sequentially selects a plurality of candidates, and sends the candidate selection its l 2-1. The candidate selection unit 12 selects the candidates as 1. The candidates are narrowed down to two and sent to the multi-syllable word recognition unit 13. Here, the DP is collated and the recognition result is sent to the ζ output via the control section 9. When it is determined that the word is a 141-syllable word, the selection unit 11 sends the feature parameters stored in the storage unit 10 to the single f1'' syllable word recognition unit 14.
Monosyllabic words are recognized and the recognition results are sent to the output via the control unit 9.

(f) Detailed explanation of the invention As described above, the present invention first classifies multi-syllabic words by matching using a reduced parameter time series, and then narrows down candidates for multi-syllabic words to be matched in detail based on the classification results and confirms them by DP matching. Therefore, it is possible to provide a speech recognition system that requires short processing time and does not require mode switching.1 The effects are significant.

[Brief explanation of drawings]

The figure is a block diagram of a circuit showing one embodiment of the present invention. 1 is a preprocessing section, and 2 is a parameter extraction section. 3.5 is a switching unit, 4 is a reduced parameter time series extraction unit, 6
is the reduced parameter time series storage unit for monosyllables. 7 is a reduced parameter time series storage unit for multiple syllables, 8 is a collation determination unit, 9 is a control unit, 10 is a storage unit, 11 is a selection unit, 1
2 is a candidate selection section, 13 is a multi-syllable word recognition section, and 14 is a monosyllabic word recognition section.

Claims (1)

[Claims] In a speech recognition system that recognizes multi-syllable words and single-f, -clausal words, there is provided a means for non-uniformly sampling speech feature parameters to obtain a reduced parameter time series of fixed length °ζ, and the reduced parameter time series. The multi-syllable words selected by the screening means are recognized after narrowing down and selecting the matching candidates to be recognized using the information at the time of screening. A multi-syllable word recognition method characterized by: