Patent US6182037B1 successfully refreshed from Google Patents.
US6182037B1
Expired / Lapsed
US
⏱️ 0.0 years remaining
Speaker recognition over large population with fast and detailed matches
Assignee: Nuance Communications Inc
•
Filing: 1997-05-06
•
Granted: 2001-01-30
•
Expires: 2017-05-06
Abstract
Abstract Fast and detailed match techniques for speaker recognition are combined into a hybrid system in which speakers are associated in groups when potential confusion is detected between a speaker being enrolled and a previously enrolled speaker. Thus the detailed match techniques are invoked only at the potential onset of saturation of the fast match technique while the detailed match is facilitated by limitation of comparisons to the group and the development of speaker-dependent models which principally function to distinguish between members of a group rather than to more fully characterize each speaker. Thus storage and computational requirements are limited and fast and accurate speaker recognition can be extended over populations of speakers which would degrade or saturate fast match systems and degrade performance of detailed match systems.
Inventors
- 👤 Stephane Herman Maes
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-09 23:18
CPC Classifications
G; G10; G10L; G10L17/00; G10L17/06
Claims Analysis
Total Claims: 24
Claim 1
Independent Claim
273 words
A method for performing speaker recognition in accordance with a computer system, said computer system including a storage unit for storing a plurality of codebooks each of which corresponds to one of a plurality of speakers, said method comprising: allocating said plurality of speakers into groups of speakers; performing a text-independent, fast-match speaker recognition process in accordance with steps that include: (a) comparing an input speech signal with a portion of said plurality of codebooks stored in said storage unit, wherein step (a) includes: (1) splitting said input speech signal into a plurality of frames, (2) for each frame: (i) deriving at least one feature vector from said input speech signal, (ii) comparing the at least one feature vector in each frame with codewords in said portion of said plurality of codebooks, the codewords in each codebook being formed from feature vectors derived from test data previously input for a corresponding one of said speakers, (3) counting frames which correspond to each of said plurality of codebooks, and (4) providing an indication of how closely each of said codebooks match said input speech signal based on said frame counting step, and (b) identifying a predetermined number of said portion of said plurality of codebooks which best match said input speech signal; and performing a detailed-match speaker recognition process in accordance with steps that include: (c) identifying groups into which the speakers identified in step (b) belong; (d) comparing said input speech signal only with models corresponding to speakers in said groups identified in step (c); and (e) identifying the speaker of said input speech signal based on an outcome of comparing step (d).
Claim 2
Independent Claim
273 words
A method for performing speaker recognition in accordance with a computer system, said computer system including a storage unit for storing a plurality of codebooks each of which corresponds to one of a plurality of speakers, said method comprising: allocating said plurality of speakers into groups of speakers; performing a text-independent, fast-match speaker recognition process in accordance with steps that include: (a) comparing an input speech signal with a portion of said plurality of codebooks stored in said storage unit, wherein step (a) includes: (1) splitting said input speech signal into a plurality of frames, (2) for each frame: (i) deriving at least one feature vector from said input speech signal, (ii) comparing the at least one feature vector in each frame with codewords in said portion of said plurality of codebooks, the codewords in each codebook being formed from feature vectors derived from test data previously input for a corresponding one of said speakers, (3) counting frames which correspond to each of said plurality of codebooks, and (4) providing an indication of how closely each of said codebooks match said input speech signal based on said frame counting step, and (b) identifying a predetermined number of said portion of said plurality of codebooks which best match said input speech signal; and performing a detailed-match speaker recognition process in accordance with steps that include: (c) identifying groups into which the speakers identified in step (b) belong; (d) comparing said input speech signal only with models corresponding to speakers in said groups identified in step (c); and (e) identifying the speaker of said input speech signal based on an outcome of comparing step (d).
Claim 3
Dependent Claim
44 words
A method as recited in claim 1 , further comprising: verifying, at a time subsequent to identification of said speaker in step (e), that the speaker of said input speech signal is likely to be a speaker previously identified by said identifying step (e).
Claim 4
Dependent Claim
35 words
A method as recited in claim 2 , wherein said verification step includes the step of determining that the speaker of said input speech signal is in a common group with said previously identified speaker.
Claim 5
Dependent Claim
22 words
A method as recited in claim 3 , including the further step of repeating said identifying step to complete said verification step.
Claim 6
Dependent Claim
72 words
A method as recited in claim 1 , wherein said allecating step includes the further steps of clustering feature vectors representing utterances of a new speaker during enrollment to form a codebook for said new speaker, decoding said feature vectors against codebooks for previously enrolled speakers, associating said new speaker and a previously enrolled speaker in a group when potential confusion is detected between said new speaker and said previously enrolled speaker.
Claim 7
Dependent Claim
26 words
A method as recited in claim 5 , including the further step of developing a speaker-dependent model capable of distinguishing between speakers associated in a group.
Claim 8
Dependent Claim
47 words
A method as recited in claim 1 , wherein said step of identifying said speaker includes the further step of determining at least two candidate speakers for recognition, and selecting a more likely of the candidate speakers if the candidate speakers are not associated in a group.
Claim 9
Dependent Claim
66 words
A method as recited in claim 1 , wherein said step of identifying said speaker includes the further step of determining at least three candidate speakers for recognition, at least two and less than all of said at least three candidate speakers being associated in a group, and eliminating a candidate speaker of said at least three candidate speakers which is not associated in said group.
Claim 10
Dependent Claim
23 words
A method as recited in claim 1 , wherein said first speaker recognition process and said second speaker recognition process are performed hierarchically.
Claim 11
Dependent Claim
17 words
A method as recited in claim 9 , wherein said identifying step is performed by vector clustering.
Claim 12
Dependent Claim
44 words
A method as recited in claim 1 , including the further step of verifying that the speaker of a further input speech signal is likely to be a speaker identified by said identifying step (e) by repeating at least said second speaker recognition process.
Claim 13
Dependent Claim
13 words
A method as recited in claim 1 , wherein identifying step is text-dependent.
Claim 14
Dependent Claim
37 words
A method as recited in claim 1 , including the further step of matching the input speed signal against speaker-dependent models within said input speed signal against speaker-dependent identity of a speaker of said input speed signal.
Claim 15
Dependent Claim
40 words
A method as recited in claim 13 , wherein said second speech recognition process includes comparing said input speech signal using a trained, speaker-dependent model and identifying a speaker as the speaker on whom the most likely model is trained.
Claim 16
Dependent Claim
19 words
A method as recited in claim 1 , wherein said models are one of speaker-dependent models or cohort models.
Claim 17
Dependent Claim
16 words
A method as recited in claim 1 , wherein said second speaker recognition process is text-dependent.
Claim 18
Dependent Claim
16 words
A method as recited in claim 1 , wherein said second speaker recognition process is text-independent.
Claim 19
Dependent Claim
32 words
A method as recited in claim 1 , further comprising: performing a verification step to determine whether at least one subsequent input speech signal is from the speaker identified in step (e).
Claim 20
Dependent Claim
42 words
A method as recited in claim 18 , further comprising: determining that said verification step failed when said performing step identifies that said subsequent input speech signal is from a speaker in a different group from the speaker identified in step (e).
Claim 21
Dependent Claim
31 words
A method as recited in claim 19 , further comprising: preventing the speaker in said different group from accessing a secure system if said verification step is determined to have failed.
Claim 22
Dependent Claim
46 words
A method as recited in claim 1 , wherein if the speakers identified in step (b) belong to a same group, said second speaker recognition process includes: using speaker-dependent models of all cohorts of the same group to determine the speaker of said input speech signal.
Claim 23
Dependent Claim
88 words
A method as recited in claim 1 , wherein said allocating step includes: enrolling at least one of said plurality of speakers in accordance with steps that include: comparing a test speech signal of said speaker to be enrolled with codebooks of already-enrolled speakers, identifying codebooks of already-enrolled speakers that are a close match with the test speech signal of said speaker to be enrolled, and defining a first class of similar speakers containing said speaker to be enrolled and said already-enrolled speaker that are a close match.
Claim 24
Dependent Claim
43 words
A method as recited in claim 22 , wherein said allocating step includes: adapting at least one of a model developed for said speaker to be enrolled and models of said already-enrolled speakers that are a close match to increase a discrimination therebetween.
References Cited (Backward Citations)
US3673331A
Cited
US4363102A
Cited
US4716593A
Cited
US4720863A
Cited
JPS59111699A
Cited
JPS612599A
Cited
US5241649A
Cited
US5347595A
Cited
US4947436A
Cited
US4827518A
Cited
US5121428A
Cited
US5384833A
Cited
US5216720A
Cited
US5073939A
Cited
US5189727A
Cited
JPH0415700A
Cited
US5274695A
Cited
US5167004A
Cited
US5271088A
Cited
US5608840A
Cited
US5682464A
Cited
US5339385A
Cited
US5412738A
Cited
US5675704A
Cited
US5537488A
Cited
US5689616A
Cited
US5522012A
Cited
US5414755A
Cited
US5666466A
Cited
US5895447A
Cited
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 2004-01-30 | 2004-07-31 | 2005-01-30 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2008-01-30 | 2008-07-31 | 2009-01-30 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2012-01-30 | 2012-07-31 | 2013-01-30 | $7700.00 | Window Closed / Lapsed |