EP0146434A1
Expired / Lapsed
EP
⏱️ 0.0 years remaining
A speaker independent speech recognition process
Assignee: Texas Instruments France SAS
•
Filing: 1984-11-07
•
Granted: —
•
Expires: 2004-11-07
Abstract
Abstract According to this process, a speech signal is analyzed in a vector quantizer (1) in which the accoustic parameters are calculated for each interval of time of a predetermined value and are compared with each spectral template contained in a template dictionary (2) utilizing a distance calculation, the sequence obtained at the output of the vector quantizer (1) is compared with a word lexicon (5) memorized in a phonetic form utilizing phonetic distribution tables (3) associated with each template.
Inventors
- 👤 Gérard Benbassat
Bibliographic Details
Application Number
—
Jurisdiction
EP
Data Source
google_patents
Last Synchronized
2026-11-01 10:27
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/06; G10L15/065; G10L15/07; G10L19/00; G10L19/02; G10L19/032; G10L19/038
Claims Analysis
Total Claims: 13
Claim 1
Independent Claim
82 words
1 - A speaker independent speech recognition method comprising dividing the speech to be analyzed into phonetic units and in comparing these phonetic units with the phonemes contained in a dictionarycharacterized in that the identification of the phonetic units consists in dividing the acoustic space into domains ,in rendering each of the domains of the acoustical space representative of a plurality of phonetic units by obtaining during a training period a probability of association of a phonetic unit to a corresponding domain.
Claim 2
Dependent Claim
82 words
1 - A speaker independent speech recognition method comprising dividing the speech to be analyzed into phonetic units and in comparing these phonetic units with the phonemes contained in a dictionarycharacterized in that the identification of the phonetic units consists in dividing the acoustic space into domains ,in rendering each of the domains of the acoustical space representative of a plurality of phonetic units by obtaining during a training period a probability of association of a phonetic unit to a corresponding domain.
Claim 3
Dependent Claim
51 words
2 - A method according to claim 1, characterized in that the maximum distance utilized for associating a point with a domain of acoustic space and consequently . to a certain number of phonetic units is. rendered as short as necessary by the choice of the number of sufficiently large domains.
Claim 4
Dependent Claim
57 words
3 - A method according to one of the claims 1 or 2, characterized in that the speech signal is analyzed in a vector quantizer (1) in which the acoustic parameters are calculated for each interval of time of a predetermined value and are compared with each template contained in a dictionary (2) utilizing a distance calculation.
Claim 5
Dependent Claim
20 words
4 - A method according to claim 3,characterized in that said distance calculation is a calculation of the euclidian distance.
Claim 6
Dependent Claim
48 words
5 - A method according to one of the claims 3 or 4,characterized in that the sequence obtained at the output of the vector quantizer (1) is compared with a lexicon fo words (5) stored in phonetic form utilizing tables (3T20) of phonetic frequencies associated with each template.
Claim 7
Dependent Claim
39 words
6 - A method according to claim 5,characterized in that a dynamic programming algorithm is utilized for calculating the probability of correspondence between the sequence obtained at the output of the vector quantizer and each word of the lexicon.
Claim 8
Dependent Claim
25 words
7 - A method according to claim 6 characterized in that the calculation of the probabilities is based on the utilization of phonetic frequency tables.
Claim 9
Dependent Claim
112 words
8 - A method according to one of the claims 1 through 7,characterized in that there is selected in a training set (8) of spectral templates (10) of which the distance from their closest neighbour in a dictionary of spectracl templates (9) is larger than a threshold,there is grouped the spectral templates (10) of the training set into classes (11) as a function of their nearest neighbour in the dictionary and there is selected,for insertion into the dictionary (9) the center of gravity (12) of each class, this process being repeated until the average distance is less than a certain threshold or until its variation becomes lower than a very low value.
Claim 10
Dependent Claim
66 words
9 - A method according to any one of the claims 1 through 8,characterized in that the training _step includes creating an analysis index (25)and a marking index (29) from the speech of the training set spoken by a predetermined number of speakers having different accents and tones and converted into digital signals, and from the analysis index and the marking index,to create frequencey tables (33).
Claim 11
Dependent Claim
76 words
10 - A method according to one of the preceding claims,characterized in that the constitution of the analysis index is assured by a coding (22) of the digital signals resulting from the conversion of the speech signals of the training set and in analysing the digital signals by an analysis by linear prediction,while the creation of the marking index (29) is effected by submitting the data of the analysis index (25) to a phonetic marking operation.
Claim 12
Dependent Claim
52 words
11 - A method according to claim 9,characterized in in that said phonetic marking operation is effected utilizing a listener (27) and a simultaneous spectral and simultaneous temporal graphic representation (28) of the digital speech signal of the training so as to determine the limits of the stable portions of the phonemes.
Claim 13
Dependent Claim
62 words
12 - A method according to one of the claims 8 to 10,characterized in that the frequency tables (33) are obtained in performing optimal selection of the spectral templates from the contents of the marking and analysis indexes (25,29) by placing the optimal templates in a dictionary index (31) and in calculating the frequencies from the contents of the dictionary index (31).
References Cited (Backward Citations)
No citation records indexed.
USPTO Annuity & Maintenance Windows
Maintenance schedule is calculated for granted US utility patents.