University Homepage School Homepage School Contacts School Search

Lecture 15 - Sequence discriminative training; Robust speech recognition

This lecture mainly covered sequence discriminative training for GMM and neural network systems. In the last part we discussed robust speech recognition

Sequence discriminative training

• Maximum likelihood vs discriminative training
• Framewise and sequence objective functions
• Maximum mutual information estimation - expressed as a numerator part (clamped to a reference word sequence) and denominator part (free, evaluating all possible word sequences)
• Use of lattice in MMIE training, for computing the denominator
• Minimum phone error criterion - explicitly weighting with a phone error rate term
• Applying sequence discriminative training to neural network acoustic models
• Results comparing ML and discriminatively trained GMM systems and framewise and sequence trained NN systems on Switchboard. Sequence discriminative training gives 10-15% (relative) reduction in word error rate.

Robust speech recognition

• Causes of additive noise and reverberation requiring robust speech recognition
• Measuring noise level using SNR
• Feature normalisation using cepstral mean and variance normalisation
• Feature compensation using spectral subtraction
• Multi-condition training - train with speech data in a variety of noise conditions

---

This page maintained by Steve Renals.
Last updated: 2017/03/28 12:58:24UTC