Loading...

A novel hybrid HMM/ANN structure for discriminative training in speech recognition

Gholampour, I ; Sharif University of Technology | 2000

59 Viewed
  1. Type of Document: Article
  2. Publisher: Sharif University of Technology , 2000
  3. Abstract:
  4. In this paper, a new formulation for discriminative training of HMMs is introduced as a solution to several speech recognition problems. This formulation uses a properly trained MLP in a simple interconnection with HMMs called "Cascade HMM/ANN Hybrid". The training algorithm has simple realization in comparison with other discriminative training for HMMs such as MDI and MMI. Also a rigid mathematical proof of its convergence has been presented. No significant increase in computational requirements is needed in recognition phase and the recognition task can still be performed in real-time. This structure has been employed in some isolated and continuous speaker-independent speech recognition experiments, including isolated word recognition in a noisy room and public telephone environments as well as continuous speech phoneme recognition in a low noise environment. The experiments demonstrate that the cascade structure is capable of better discrimination of closely pronounced speech units than classical HMM-based recognizers. It also resulted in significant improvement of recognition rate, about 10% for isolated word and 8% for continuous speech phoneme recognition. Moreover, in telephony isolated word recognition experiments, this structure is more robust to endpoints of the words in the presence of background noise, particularly in mobile telephones. Both theoretical and experimental results are presented in this paper
  5. Keywords:
  6. Algorithm ; Speech discrimination ; Mathematical analysis ; Noise ; Artificial neural network
  7. Source: Scientia Iranica ; Volume 7, Issue 3-4 , 2000 , Pages 186-196 ; 10263098 (ISSN)
  8. URL: http://scientiairanica.sharif.edu/article_2791.html