Summary

This PhD thesis investigates decoding strategies and discriminative training methods to improve the speed and accuracy of large-vocabulary continuous speech recognition systems. It explains practical trade-offs in decoder design, search/pruning techniques, and sequence-level discriminative training for acoustic models, with guidance on improving robustness in varied conditions.

Key Takeaways

• Apply sequence-discriminative training to acoustic models to reduce recognition errors compared with purely maximum-likelihood methods.
• Tune decoding parameters (beam width, pruning thresholds, insertion penalties) to balance recognition speed and search errors.
• Use lattice generation and rescoring strategies to enable efficient post-decoding model improvements without full re-decoding.
• Combine robust feature processing with discriminative criteria to improve ASR performance in noisy or mismatched environments.

Who Should Read This

Advanced researchers and engineers in speech and audio processing or machine learning working on acoustic modeling, decoder implementation, or system-level ASR performance optimization.

Still RelevantAdvanced

Topics