EP0430615A2
Expired / Lapsed
EP
⏱️ 0.0 years remaining
Speech recognition system
Assignee: Toshiba Corp
•
Filing: 1990-11-26
•
Granted: —
•
Expires: 2010-11-26
Abstract
Abstract A speech recognition system comprising a speech analysis unit (2), two speech recognition units (7, 10), a segmentation unit (8), a feature vector extraction unit (9), and a recognition result processing unit (11). The analysis unit (2) analyzes an input speech containing segments, to obtain feature parameters of the segments. The first recognition unit (7) has a first dictionary containing reference patterns of recognition categories, obtains first feature vectors of the segments from the the feature parameters obtained by the analysis unit, without detecting boundaries of the segments, and collates the first feature vectors with the reference patterns stored in the first dictionary, to obtain first similarities of the recognition categories. The segmentation unit (8) detects the boundaries of the segments from the first feature vectors. The vector extracting unit (9) extracts second feature vectors from the feature parameters, based on the segment boundaries detected by the segmentation unit. The second recognition unit (10) has a second recognition dictionary storing reference patterns of the recognition categories. It collates the second feature vectors extracted by the vector extraction unit, with the reference patterns stored in the second dictionary, thereby to obtain second similarities of the recognition categories. The recognition-result processing unit obtains speech recognition results from the first similarities and the second similarities.
Inventors
- 👤 Yoichi C/O Intellectual Property Div Takebayashi
- 👤 Hiroyuki C/O Intellectual Property Div. Tsuboi
- 👤 Hiroshi C/O Intellectual Property Div. Kanazawa
Bibliographic Details
Application Number
—
Jurisdiction
EP
Data Source
google_patents
Last Synchronized
2026-10-28 04:04
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/04; G10L15/05; G10L15/02; G10L15/08; G10L2015/088; G10L25/00; G10L25/27
Claims Analysis
Total Claims: 16
Claim 1
Independent Claim
235 words
A speech recognition system characterized by comprising: speech analysis means (2) for analyzing an input speech having a plurality of segments, to obtain time-series data of feature parameters of the segments; first speech recognition means (7) having a first recognition dictionary containing reference patterns of recognition categories, and designed to obtain time-series data of first feature vectors of the segments from the time-series data obtained by said speech analysis means, without detecting boundaries of the segments, and also to collate the first feature vectors with the reference patterns stored in the first recognition dictionary, to obtain time-series data of first similarities of the recognition categories; segmentation means (8) for detecting the boundaries of the segments from the first feature vectors; feature vector extracting means (9) for extracting second feature vectors from the feature parameters, in accordance with the boundaries of the segments which have been detected by said segmentation means; second speech recognition means (10) having a second recognition dictionary storing reference patterns of the recognition categories, and designed to collate the second feature vectors extracted by said feature vector extraction means, with the reference patterns stored in said second recognition dictionary, to obtain time-series data of second similarities of the recognition categories; and recognition-result processing section (11) for obtaining speech recognition results from the first similarities obtained by said first speech recognition section and the second similarities obtained by said second speech recognition section.
Claim 2
Dependent Claim
25 words
The system according to claim 1, characterized in that the first feature vectors are of an order lower than that of the second feature vectors.
Claim 3
Dependent Claim
47 words
The system according to claim 1, characterized in that said first speech recognition means (7) includes means for collating the first feature vectors with the reference patterns stored in said first recognition dictionary by means of word spotting method, to obtain the time-series data of first similarities.
Claim 4
Dependent Claim
41 words
The system according to claim 1, characterized in that said segmentation means (8) includes means for detecting the boundaries of the segments of the input speech in accordance with the time-series data of similarities obtained by said first speech recognition means.
Claim 5
Dependent Claim
43 words
The system according to claim 1, characterized in that said first and second speech recognition means (7,10) calculate, in the same way, the similarities between each of the feature vectors and the reference patterns, to collate the feature vectors with the reference patterns.
Claim 6
Dependent Claim
250 words
A speech recognition system characterized by comprising: speech analysis means (2) for analyzing an input speech having a plurality of segments, to obtain time-series data of feature parameters of the segments; first speech recognition means (7) having a first recognition dictionary containing reference patterns of recognition categories, and designed to obtain time-series data of first feature vectors of the segments from the time-series data obtained by said speech analysis means, without detecting boundaries of the segments, and also to collate the first feature vectors with the reference patterns stored in the first recognition dictionary, to obtain time-series data of first similarities of the recognition categories; segmentation means (8) for detecting the boundaries of the segments from the first feature vectors; feature vector extracting means (9) for extracting second feature vectors from the feature parameters, in accordance with the boundaries of the segments which have been detected by said segmentation means; second speech recognition means (10) having a second recognition dictionary storing reference patterns of the recognition categories, and designed to collate the second feature vectors extracted by said feature vector extraction means, with the reference patterns stored in second recognition dictionary, to obtain time-series data of second similarities of the recognition categories; recognition-result processing section (11) for obtaining speech recognition results from the first similarities obtained by said first speech recognition section and the second similarities obtained by said second speech recognition section; and speech pattern learning means (21,25) for updating the contents of said first and second recognition dictionaries.
Claim 7
Dependent Claim
112 words
The system according to claim 6, characterized in that said speech pattern learning means includes means (23) for processing speech data representing the input speech, and forming modified speech data, means (24) for detecting start and end points for the segments of the input speech, in accordance with the time-series data of similarities of the recognition categories which have been obtained through word spotting method performed by said first speech recognition means, means (2) for analyzing the modified speech data, to obtain time-series data of feature parameters, and means for extracting the first and second feature parameters from the time-series data of feature parameters in accordance with the start and end points.
Claim 8
Dependent Claim
29 words
The system according to claim 7, characterized in that said speech pattern learning means includes means (23) for adding noise to the speech data, to modify the speech data.
Claim 9
Independent Claim
179 words
A method of recognizing an input speech, characterized by comprising: a first step of analyzing an input speech having a plurality of segments, to obtain time-series data of feature parameters of the segments; a second step of obtaining time-series data of first feature vectors of the segments from the time-series data obtained by said speech analysis means, without detecting boundaries of the segments, and collating the first feature vectors with reference patterns stored in a first recognition dictionary, to obtain time-series data of first similarities of the recognition categories; a third step of detecting the boundaries of the segments from the first feature vectors, to distinguish the segments from one another; a fourth step of extracting second feature vectors from the feature parameters, in accordance with the boundaries of the segments detected; a fifth step of collating the second feature vectors extracted, with reference patterns stored in a second recognition dictionary, to obtain time-series data of second similarities of the recognition categories; and a sixth step of obtaining speech recognition results from the first similarities and the second similarities.
Claim 10
Dependent Claim
25 words
The method according to claim 9, characterized in that the first feature vectors are of an order lower than that of the second feature vectors.
Claim 11
Dependent Claim
45 words
The method according to claim 9, characterized in that said second step includes a step of collating the first feature vectors with the reference patterns stored in the first recognition dictionary by means of word spotting method, to obtain the time-series data of first similarities.
Claim 12
Dependent Claim
39 words
The method according to claim 9, characterized in that said third step includes a step of detecting the boundaries of the segments of the input speech in accordance with the time-series data of similarities obtained by said second step.
Claim 13
Dependent Claim
44 words
The method according to claim 9, characterized in that in said second step and said fifth step, the similarities between each of the feature vectors and the reference patterns are calculated in the same way, to collate the feature vectors with the reference patterns.
Claim 14
Dependent Claim
24 words
The method according to claim 9, characterized by further comprising a seventh step of updating the contents of said first and second recognition dictionaries.
Claim 15
Dependent Claim
111 words
The method according to claim 14, characterized in that said seventh step includes a step of processing speech data representing the input speech, to form modified speech data, a step of detecting start and end points for the segments of the input speech, in accordance with the time-series data of similarities of the recognition categories which have been obtained through word spotting method performed by said first speech recognition means, a step of analyzing the modified speech data, to obtain time-series data of feature parameters, and a step of extracting the first and second feature parameters from the time-series data of feature parameters in accordance with the start and end points.
Claim 16
Dependent Claim
27 words
The method according to claim 15, characterized in that said seventh step is a step of adding noise to the speech data, to modify the speech data.
References Cited (Backward Citations)
No citation records indexed.
USPTO Annuity & Maintenance Windows
Maintenance schedule is calculated for granted US utility patents.