Lattice-free State-level Minimum Bayes Risk Training of Acoustic Models

By Naoyuki Kanda, Yusuke Fujita, Kenji Nagamatsu. 2018

29 citations recorded by the source corpus.

Lattice-free maximum mutual information (LF-MMI) training, which enables MMI-based acoustic model training without any lattice generation procedure, has recently been proposed. Although LF-MMI showed high accuracy in many tasks, its MMI criterion does not necessarily maximize the speech recognition accuracy. In this work, we propose a lattice-free state-level minimum Bayes risk training (LF-sMBR), which maximizes state-level expected accuracy without relying on a lattice generation procedure. As is the case with the LF-MMI, LF-sMBR avoids redundant lattice generation by exploiting forward-backward calculation on phone N-gram space, which enables a much simpler and faster training based on an sMBR criterion. We found that special care for silence phones was essential for improving the accuracy by LF-sMBR. In our experiments on the AMI, CSJ, and Librispeech corpora, LF-sMBR achieved small but consistent improvements over LF-MMI AMs, showing state-of-the-art results for each test set.

Subfields: Artificial Intelligence, Natural Language Processing, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

In Signal Processing

Citations in the corpus

RWTH ASR Systems for LibriSpeech: Hybrid vs Attention - w/o Data Augmentation (2019, methodology): “Previous results on LibriSpeech using hybrid models are presented in [3, 4].”
Acoustic Modeling for Distant Multi-talker Speech Recognition with Single- and Multi-channel Branches (2019, methodology): “, LibriSpeech in [6] and [7]) the level of human transcribers.”