US5950157A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Method for establishing handset-dependent normalizing models for speaker recognition
Assignee: SRI International Inc
•
Filing: 1997-04-18
•
Granted: 1999-09-07
•
Expires: 2017-04-18
Abstract
Abstract Adverse effects of type mismatch between acoustic input devices used during testing and during training in machine-based recognition of the source of acoustic phenomena are minimized. A normalizing model is matched to a source model based, or dependent, upon an acoustic input device whose transfer characteristics color acoustic characteristics of a source as represented in the source model. An application of the present invention is to speaker recognition, i.e., recognition of the identity of a speaker by the speaker's voice.
Inventors
- 👤 Larry P. Heck
- 👤 Mitchel Weintraub
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-09 21:17
CPC Classifications
G; G10; G10L; G10L17/00; G10L17/20; G10L15/00; G10L15/20
Claims Analysis
Total Claims: 40
Claim 1
Independent Claim
103 words
A method for establishing a normalizing model suitable for use with a speaker model to normalize said speaker model, said speaker model for modeling voice characteristics of a specific individual, said speaker model and said normalizing model for use in recognizing identity of a speaker, the method comprising the steps of: identifying at least one type of acoustic input device whose transfer characteristics color said voice characteristics as modeled by said speaker model; and determining, with reference to said at least one acoustic input device type, a normalizing model comprising representation of transfer characteristics of said at least one acoustic input device type.
Claim 2
Independent Claim
103 words
A method for establishing a normalizing model suitable for use with a speaker model to normalize said speaker model, said speaker model for modeling voice characteristics of a specific individual, said speaker model and said normalizing model for use in recognizing identity of a speaker, the method comprising the steps of: identifying at least one type of acoustic input device whose transfer characteristics color said voice characteristics as modeled by said speaker model; and determining, with reference to said at least one acoustic input device type, a normalizing model comprising representation of transfer characteristics of said at least one acoustic input device type.
Claim 3
Dependent Claim
72 words
The method according to claim 1, wherein the normalizing model determining step comprises determining a composite model as said normalizing model, said composite model comprising representation of voice characteristics of a plurality of speakers, said composite model obtained from training a seed model using normalizing training data from said plurality of speakers, said normalizing training data including data processed using acoustic input devices of said at least one acoustic input device type.
Claim 4
Dependent Claim
53 words
The method according to claim 2, wherein the normalizing model determining step comprises determining said composite model by training a Gaussian mixture seed model using normalizing training data from said plurality of speakers, said normalizing training data including data processed using acoustic input devices of said at least one acoustic input device type.
Claim 5
Dependent Claim
43 words
The method according to claim 1, wherein the normalizing model determining step comprises determining a model, as said normalization model, that comprises representation of transfer characteristics of types of acoustic input devices only from among said at least one acoustic input device type.
Claim 6
Dependent Claim
42 words
The method according to claim 4, wherein the normalizing model determining step comprises training a seed model using normalizing training data, including data processed using acoustic input devices of said at least one acoustic input device type, to obtain said normalizing model.
Claim 7
Dependent Claim
94 words
The method according to claim 4, wherein said at least one identified type includes a first identified type and a different second identified type; wherein the method further comprises the step of ascertaining that said first identified type of acoustic input device was used more heavily in creating said speaker model than was said second identified type; and wherein the normalizing model determining step comprises determining a model, as said normalization model, that comprises stronger representation of transfer characteristics of said first identified type than representation of transfer characteristics of said second identified type.
Claim 8
Dependent Claim
158 words
The method according to claim 1, wherein the identifying step comprises the steps of: accepting a portion of speech data used in creating said speaker model; computing a first score for said portion according to a first acoustic model that was built to comprise representation of transfer characteristics of a first type of acoustic input device, said first acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said first device type; computing a second score for said portion according to a second acoustic model that was built to comprise representation of transfer characteristics of a second type of acoustic input device, said second acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said second device type; and ascertaining that said first device type was used in creating said speaker model if said first score exceeds said second score.
Claim 9
Dependent Claim
34 words
The method according to claim 1, wherein the type identifying step comprises identifying all types of acoustic input device used in creating said speaker model as said at least one acoustic input device type.
Claim 10
Dependent Claim
43 words
The method according to claim 8, wherein the normalizing model determining step comprises determining a model, as said normalization model, that comprises representation of transfer characteristics of types of acoustic input devices only from among said at least one acoustic input device type.
Claim 11
Dependent Claim
100 words
A method for recognizing identity of a test speaker from input speech data from said test speaker, the method employing a normalizing model suitable for use with a speaker model comprising representation of voice characteristics of a reference speaker, said normalizing model being established by the method according to claim 9, the method for recognizing comprising the steps of: accepting said input speech data; computing a normalized score for said input speech data according to the speaker model and the normalizing model; and deciding, based on said normalized score, whether said input speech data was derived from said reference speaker.
Claim 12
Dependent Claim
26 words
The method according to claim 1, wherein the normalizing model determining step comprises selecting a model, as said normalizing model, from a preexisting library of models.
Claim 13
Dependent Claim
13 words
A normalizing model established in accordance with the method according to claim 1.
Claim 14
Independent Claim
148 words
A method for recognizing identity of a test speaker from input speech data from said test speaker, the method comprising the steps of: accepting said input speech data, wherein said input speech data was processed using a test acoustic input device, wherein type of said acoustic input device is not known for the purpose of recognizing said identity; computing a normalized score for said input speech data according to a speaker model and a normalizing model, said speaker model comprising representation of voice characteristics of a reference speaker and representation of transfer characteristics of a training type of acoustic input device, said normalizing model comprising representation of acoustic characteristics of sounds not spoken by said reference speaker and representation of transfer characteristics of said training type of acoustic input device; and deciding, based on said normalized score, whether said input speech data was derived from said reference speaker.
Claim 15
Independent Claim
148 words
A method for recognizing identity of a test speaker from input speech data from said test speaker, the method comprising the steps of: accepting said input speech data, wherein said input speech data was processed using a test acoustic input device, wherein type of said acoustic input device is not known for the purpose of recognizing said identity; computing a normalized score for said input speech data according to a speaker model and a normalizing model, said speaker model comprising representation of voice characteristics of a reference speaker and representation of transfer characteristics of a training type of acoustic input device, said normalizing model comprising representation of acoustic characteristics of sounds not spoken by said reference speaker and representation of transfer characteristics of said training type of acoustic input device; and deciding, based on said normalized score, whether said input speech data was derived from said reference speaker.
Claim 16
Dependent Claim
76 words
The method according to claim 13 wherein said normalizing model was established by the steps of: identifying at least one type of acoustic input device used in creating said speaker model; and determining, with reference to said at least one type, said normalizing model comprising representation of transfer characteristics of said at least one acoustic input device type, wherein said training acoustic input device type is from among said at least one acoustic input device type.
Claim 17
Independent Claim
195 words
A method for building a speaker recognition system comprising the steps of: accepting acoustic speaker training data, the speaker training data being derived from a speaker; training a seed model using said speaker training data to establish a speaker model for the speaker; and identifying at least one type of acoustic input device used in deriving said speaker training data, wherein the identifying step comprises the steps of: computing a first score for a portion of said speaker training data according to a first acoustic model that was built to comprise representation of transfer characteristics of a first type of acoustic input device, said first acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said first device type; computing a second score for said speaker training data portion according to a second acoustic model that was built to comprise representation of transfer characteristics of a second type of acoustic input device, said second acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said second device type; and comparing said first score to said second score.
Claim 18
Independent Claim
195 words
A method for building a speaker recognition system comprising the steps of: accepting acoustic speaker training data, the speaker training data being derived from a speaker; training a seed model using said speaker training data to establish a speaker model for the speaker; and identifying at least one type of acoustic input device used in deriving said speaker training data, wherein the identifying step comprises the steps of: computing a first score for a portion of said speaker training data according to a first acoustic model that was built to comprise representation of transfer characteristics of a first type of acoustic input device, said first acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said first device type; computing a second score for said speaker training data portion according to a second acoustic model that was built to comprise representation of transfer characteristics of a second type of acoustic input device, said second acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said second device type; and comparing said first score to said second score.
Claim 19
Independent Claim
89 words
A method for building a speaker recognition system comprising the steps of: accepting acoustic speaker training data, the speaker training data being derived from a speaker; training a seed model using said speaker training data to establish a speaker model for the speaker; identifying at least one type of acoustic input device used in deriving said speaker training data; and determining, with reference to said at least one acoustic input device type, a normalizing model comprising representation of transfer characteristics of said at least one acoustic input device type.
Claim 20
Independent Claim
89 words
A method for building a speaker recognition system comprising the steps of: accepting acoustic speaker training data, the speaker training data being derived from a speaker; training a seed model using said speaker training data to establish a speaker model for the speaker; identifying at least one type of acoustic input device used in deriving said speaker training data; and determining, with reference to said at least one acoustic input device type, a normalizing model comprising representation of transfer characteristics of said at least one acoustic input device type.
Claim 21
Independent Claim
168 words
A method for identifying from input speech data a device type, called the identified device type, of an acoustic input device used in processing said speech data, the method comprising the steps of: accepting said speech data; computing a first score for said speech data according to a first acoustic model that was built to comprise representation of transfer characteristics of a first type of acoustic input device, said first acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said first device type; computing a second score for said speech data according to a second acoustic model that was built to comprise representation of transfer characteristics of a second type of acoustic input device, said second acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said second device type; and providing said first device type as said identified device type if said first score exceeds said second score.
Claim 22
Independent Claim
168 words
A method for identifying from input speech data a device type, called the identified device type, of an acoustic input device used in processing said speech data, the method comprising the steps of: accepting said speech data; computing a first score for said speech data according to a first acoustic model that was built to comprise representation of transfer characteristics of a first type of acoustic input device, said first acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said first device type; computing a second score for said speech data according to a second acoustic model that was built to comprise representation of transfer characteristics of a second type of acoustic input device, said second acoustic model formed by training a seed model using training acoustic data processed using an acoustic input device of said second device type; and providing said first device type as said identified device type if said first score exceeds said second score.
Claim 23
Dependent Claim
59 words
The method according to claim 17, wherein the first score computing step comprises computing said first score according to a Gaussian mixture model, said Gaussian mixture model being said first acoustic model; and wherein the second score computing step comprises computing said second score according to a Gaussian mixture model, said Gaussian mixture model being said second acoustic model.
Claim 24
Dependent Claim
174 words
A speaker recognition system for recognizing identity of a test speaker from input speech data from said test speaker, wherein said input speech data was processed using an acoustic input device of a test type from among expected test types, the system comprising: a speaker model comprising representation of voice characteristics of a reference speaker and representation of transfer characteristics of a type of acoustic input device herein called the represented type, said speaker model comprising representation of less than all of said expected test types; a normalizing model comprising representation of acoustic characteristics of sounds not spoken by said reference speaker and representation of transfer characteristics of said represented acoustic input device type; and a scoring module coupled to said speaker model and said normalizing model and coupled to accept said input speech data, the scoring module for computing from said input speech data a normalized score using said speaker model and said normalizing model, said normalized score indicating similarity between said input speech data and said voice characteristics of said reference speaker.
Claim 25
Dependent Claim
174 words
A speaker recognition system for recognizing identity of a test speaker from input speech data from said test speaker, wherein said input speech data was processed using an acoustic input device of a test type from among expected test types, the system comprising: a speaker model comprising representation of voice characteristics of a reference speaker and representation of transfer characteristics of a type of acoustic input device herein called the represented type, said speaker model comprising representation of less than all of said expected test types; a normalizing model comprising representation of acoustic characteristics of sounds not spoken by said reference speaker and representation of transfer characteristics of said represented acoustic input device type; and a scoring module coupled to said speaker model and said normalizing model and coupled to accept said input speech data, the scoring module for computing from said input speech data a normalized score using said speaker model and said normalizing model, said normalized score indicating similarity between said input speech data and said voice characteristics of said reference speaker.
Claim 26
Dependent Claim
33 words
The system according to claim 19, wherein said normalizing model comprises representation of transfer characteristics of acoustic input device types only from among types whose transfer characteristics are represented in said speaker model.
Claim 27
Dependent Claim
92 words
The system according to claim 19, wherein: said speaker model comprises representation of transfer characteristics of, in addition to said represented acoustic input device type herein called the first type, a second type of acoustic input device, said representation of transfer characteristics of said second type being stronger than said representation of said first type; and said normalizing model comprises representation of transfer characteristics of said first and second types of acoustic input devices, said representation of transfer characteristics of said second type being stronger than said representation of said first type.
Claim 28
Dependent Claim
162 words
A speaker recognition system for recognizing identity of a test speaker from input speech data from said test speaker, wherein said input speech data was processed using an acoustic input device of a test type from among expected test types, the system comprising: a speaker model comprising representation of voice characteristics of a reference speaker and representation of transfer characteristics of said expected test types; a normalizing model comprising representation of acoustic characteristics of sounds not spoken by said reference speaker and representation of transfer characteristics of a type of acoustic input device whose transfer characteristics are represented in said speaker model; and a scoring module coupled to said speaker model and said normalizing model and coupled to accept said input speech data, the scoring module for computing from said input speech data a normalized score using said speaker model and said normalizing model, said normalized score indicating similarity between said input speech data and said voice characteristics of said reference speaker.
Claim 29
Dependent Claim
162 words
A speaker recognition system for recognizing identity of a test speaker from input speech data from said test speaker, wherein said input speech data was processed using an acoustic input device of a test type from among expected test types, the system comprising: a speaker model comprising representation of voice characteristics of a reference speaker and representation of transfer characteristics of said expected test types; a normalizing model comprising representation of acoustic characteristics of sounds not spoken by said reference speaker and representation of transfer characteristics of a type of acoustic input device whose transfer characteristics are represented in said speaker model; and a scoring module coupled to said speaker model and said normalizing model and coupled to accept said input speech data, the scoring module for computing from said input speech data a normalized score using said speaker model and said normalizing model, said normalized score indicating similarity between said input speech data and said voice characteristics of said reference speaker.
Claim 30
Dependent Claim
26 words
The system according to claim 22, wherein said expected test acoustic input device types are a carbon microphone telephone type and an electret microphone telephone type.
Claim 31
Dependent Claim
20 words
The system according to claim 22, wherein said normalizing model comprises representation of transfer characteristics of all said expected types.
Claim 32
Dependent Claim
31 words
The system according to claim 22, wherein said normalizing model comprises representation of transfer characteristics of all types of acoustic input devices whose transfer characteristics are represented in said speaker model.
Claim 33
Dependent Claim
38 words
The system according to claim 22, wherein all acoustic input device types whose transfer characteristics are represented in said normalizing model are from among types of acoustic input devices whose transfer characteristics are represented in said speaker model.
Claim 34
Dependent Claim
93 words
The system according to claim 22, wherein: said speaker model comprises representation of transfer characteristics of, in addition to said represented acoustic input device type herein called the first type, a second type of acoustic input device, said representation of transfer characteristics of said second type being stronger than said representation of said first type; and wherein said normalizing model comprises representation of transfer characteristics of said first and second types of acoustic input devices, said representation of transfer characteristics of said second type being stronger than said representation of said first type.
Claim 35
Independent Claim
148 words
A system for establishing a normalizing model suitable for use with a speaker model to normalize said speaker model, said speaker model for modeling voice characteristics of a specific individual, said speaker model and said normalizing model for use in recognizing identity of a speaker, the system comprising: an input stage coupled to accept a signal indicative of at least one acoustic input device type whose transfer characteristics color said voice characteristics as modeled by said speaker model, said input stage for identifying said at least one speaker acoustic input device type from said signal and providing said at least one type; and a controller coupled to receive said at least one speaker input device type, said controller for determining, with reference to said at least one acoustic input device type, a normalizing model comprising representation of transfer characteristics of said at least one acoustic input device type.
Claim 36
Independent Claim
148 words
A system for establishing a normalizing model suitable for use with a speaker model to normalize said speaker model, said speaker model for modeling voice characteristics of a specific individual, said speaker model and said normalizing model for use in recognizing identity of a speaker, the system comprising: an input stage coupled to accept a signal indicative of at least one acoustic input device type whose transfer characteristics color said voice characteristics as modeled by said speaker model, said input stage for identifying said at least one speaker acoustic input device type from said signal and providing said at least one type; and a controller coupled to receive said at least one speaker input device type, said controller for determining, with reference to said at least one acoustic input device type, a normalizing model comprising representation of transfer characteristics of said at least one acoustic input device type.
Claim 37
Dependent Claim
165 words
The system according to claim 28, wherein said signal is a portion of training data used in establishing said speaker model, and wherein said input stage is a type detector comprising: a first input device model comprising representation of transfer characteristics of a first type of acoustic input device; a second input device model comprising representation of transfer characteristics of a second type of acoustic input device; a scoring module coupled to said first and second device models and coupled to accept said input data, said scoring module for computing from said input data a first device score using said first device model and a second device score using said second device model; and a comparator coupled to receive said first device score and said second device score, said comparator for comparing said first and second scores and if said first score is greater than said second score identifying and providing said first type as one of said at least one acoustic input device type.
Claim 38
Dependent Claim
74 words
The system according to claim 28 further comprising a normalizing model trainer for training a seed model using normalizing training data from a plurality of speakers other than said individual, said normalizing training data including data processed using an acoustic input device of said at least one acoustic input device type, said normalizing model trainer coupled to said controller, said controller providing said trained model from said normalizing model trainer as said normalizing model.
Claim 39
Dependent Claim
160 words
A type detector for identifying from input speech data a device type, called the identified device type, of an acoustic input device used in processing said speech data, the type detector comprising: a first input device model comprising representation of transfer characteristics of a first type of acoustic input device; a second input device model comprising representation of transfer characteristics of a second type of acoustic input device; a scoring module coupled to said first and second device models and coupled to accept said input speech data, said scoring module for computing from said input speech data a first device score using said first device model and a second device score using said second device model; and a comparator coupled to receive said first device score and said second device score, said comparator for comparing said first and second scores and if said first score is greater than said second score indicating said first type as said identified device type.
Claim 40
Dependent Claim
160 words
A type detector for identifying from input speech data a device type, called the identified device type, of an acoustic input device used in processing said speech data, the type detector comprising: a first input device model comprising representation of transfer characteristics of a first type of acoustic input device; a second input device model comprising representation of transfer characteristics of a second type of acoustic input device; a scoring module coupled to said first and second device models and coupled to accept said input speech data, said scoring module for computing from said input speech data a first device score using said first device model and a second device score using said second device model; and a comparator coupled to receive said first device score and said second device score, said comparator for comparing said first and second scores and if said first score is greater than said second score indicating said first type as said identified device type.
References Cited (Backward Citations)
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 2002-09-06 | 2003-03-08 | 2003-09-07 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2006-09-06 | 2007-03-08 | 2007-09-07 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2010-09-06 | 2011-03-08 | 2011-09-07 | $7700.00 | Window Closed / Lapsed |