Reconnaissance automatique de la parole continue en environnement bruité : application à des modèles stochastiques de trajectoires

Abstract : Most automatic speech recognisers perform poorly when the training and testing conditions are not matched. This dissertation describes a number of algorithms that improve the noise robustness of the VINICS continuo us speech recogniser, developed at CRIN-INRIA Lorraine and based on stochastic modeling of speech trajectories. First, we give a review of recent works dealing with noisy speech recognition. Second, we propose and compare three noisy speech recognition approaches. In the first one, a noisy speech Stochastic Trajectory Model (STM) is derived from a clean speech STM and an hidden Markov model of noise. The second approach performs an STM state-based filtering of the noisy speech parametric signal, where the estimators are derived in the logspectrum domain. In the last method, we develop an adaptation framework based on linear regression to adapt STM means to new environments. Experiments on a continuous speech recognition task on speech corrupted by different additive noises show that the adaptation using linear regression outperforms the other approaches. In the last part, we first study the noise robustness of speech parameters derived using linear discriminant analysis. The derived parameters are very efficient when the training and the testing signal-to-noise ratios are matched, but are very sensitive to signal-to-noise ratio variations. Then, we propose the use of Bayesian reestimation to reduce the mismatches between phone duration in Lombard and clean speech. On an isolated word speech recognition task, the adaptation of phone duration models significantly improve the recognition rate.
Document type :
Theses
Complete list of metadatas

https://hal.univ-lorraine.fr/tel-01748602
Contributor : Thèses Ul <>
Submitted on : Thursday, March 29, 2018 - 11:42:46 AM
Last modification on : Tuesday, January 29, 2019 - 1:21:56 AM
Long-term archiving on: Friday, September 14, 2018 - 4:43:11 AM

File

SCD_T_1995_0290_SIOHAN.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : tel-01748602, version 1

Collections

Citation

Olivier Siohan. Reconnaissance automatique de la parole continue en environnement bruité : application à des modèles stochastiques de trajectoires. Autre [cs.OH]. Université Henri Poincaré - Nancy 1, 1995. Français. ⟨NNT : 1995NAN10290⟩. ⟨tel-01748602⟩

Share

Metrics

Record views

46

Files downloads

73