Grant: Discovery Grants Program - Individual
Institut national de la recherche scientifique
$52,625.00
Improving basic methods of automatic speech recognition
FY 2012-13
Summary
The biggest problem in automatic speech recognition (ASR) is robustness - making sure that ASR models are general enough. A robust ASR system should be able to decode speech from any speaker, in any reasonable environment. In practice, speakers other than those that trained the ASR models lead to a significant decrease in ASR accuracy. As well, environmental noise (e.g., weather, other talkers, cars) and communication distortions (e.g., with mobile phones) degrade ASR performance, often severely. Even simple microphone substitution reduces accuracy. Human listeners, in contrast, often can adapt rapidly to all these difficulties, which strongly suggests that major flaws exist in current ASR schemes. While it is not necessary to directly mimic human perception when designing ASR systems, much of what we know about human speech production and perception has yet to be integrated into practical ASR. ASR can be greatly improved by better integration of structural information with the current stochastic methodology.
Over the last 25 years, the ASR field has accepted certain technical procedures as almost de facto standards. A large majority of ASR divides up speech signals into short blocks of data (called frames, updated typically every 10 ms), does spectral analysis on each frame leading to a set of 10-50 parameters, applies a hidden Markov model (HMM) to decode the speech, and uses a simple language model to help determine the output word sequence. We propose to re-evaluate some of these crticial methods, in ways that include speculative and emerging techniques. In particular, we will address several of the flaws of the popular MFCC (mel-frequency cepstral coefficient) method and of the HMM approach.
Program
Discovery Grants Program - Individual
Location
Quebec, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
914-2008
Name nom
OShaughnessy, Douglas
Department département
Centre Énergie, Matériaux et Télécommunication - Montréal
Organization id
56
Institution établissement
Institut national de la recherche scientifique
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2012
Competition year année de concours
2008
Award amount
52625
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
Discovery grants programs
Group fr
Programmes de subventions à la découverte
Committee code
334
Committee name en
Commun.,Computers & Components
Committee name fr
Communic.,ordin. et composants
Area of application code
802
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Information systems and technology
Area of application fr
Systèmes d'information et technologie de l'information
Research subject code
2700
Research subject group en
Information technology
Research subject group fr
Technologie de l'information
Research subject en
Information technology
Research subject fr
Technologie de l'information
Partie
5
Nb partie
5
Application title
Improving basic methods of automatic speech recognition
Keyword
NSERC
Application summary
The biggest problem in automatic speech recognition (ASR) is robustness - making sure that ASR models are general enough. A robust ASR system should be able to decode speech from any speaker, in any reasonable environment. In practice, speakers other than those that trained the ASR models lead to a significant decrease in ASR accuracy. As well, environmental noise (e.g., weather, other talkers, cars) and communication distortions (e.g., with mobile phones) degrade ASR performance, often severely. Even simple microphone substitution reduces accuracy. Human listeners, in contrast, often can adapt rapidly to all these difficulties, which strongly suggests that major flaws exist in current ASR schemes. While it is not necessary to directly mimic human perception when designing ASR systems, much of what we know about human speech production and perception has yet to be integrated into practical ASR. ASR can be greatly improved by better integration of structural information with the current stochastic methodology.
Over the last 25 years, the ASR field has accepted certain technical procedures as almost de facto standards. A large majority of ASR divides up speech signals into short blocks of data (called frames, updated typically every 10 ms), does spectral analysis on each frame leading to a set of 10-50 parameters, applies a hidden Markov model (HMM) to decode the speech, and uses a simple language model to help determine the output word sequence. We propose to re-evaluate some of these crticial methods, in ways that include speculative and emerging techniques. In particular, we will address several of the flaws of the popular MFCC (mel-frequency cepstral coefficient) method and of the HMM approach.