Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Investigating the Effect of Label Topology and Training Criterion on ASR Performance and Alignment Quality

Tina Raissi, Christoph Lüscher, Simon Berger, Ralf Schlüter, Hermann Ney

TL;DR

The paper tackles the comparability gap between end-to-end and modular ASR by isolating the effects of label topology and training criteria. It jointly evaluates zero-order alignment models (CTC and Posterior-HMM) and first-order label-context models (Factored Hybrid HMM and Monotonic RNN-T) under fixed front-end and downsampling conditions, across LibriSpeech 960h and Switchboard 300h. Key findings show CTC can yield competitive WER and TSE on alignments, but subsequent first-order training guided by those alignments often achieves better ASR performance, with a 40 ms frame shift delivering substantial real-time factor speedups. The study demonstrates that frame-rate and topology choices materially influence both alignment quality and production-time efficiency, offering practical guidance for designing ASR pipelines that balance accuracy and latency.

Abstract

The ongoing research scenario for automatic speech recognition (ASR) envisions a clear division between end-to-end approaches and classic modular systems. Even though a high-level comparison between the two approaches in terms of their requirements and (dis)advantages is commonly addressed, a closer comparison under similar conditions is not readily available in the literature. In this work, we present a comparison focused on the label topology and training criterion. We compare two discriminative alignment models with hidden Markov model (HMM) and connectionist temporal classification topology, and two first-order label context ASR models utilizing factored HMM and strictly monotonic recurrent neural network transducer, respectively. We use different measurements for the evaluation of the alignment quality, and compare word error rate and real time factor of our best systems. Experiments are conducted on the LibriSpeech 960h and Switchboard 300h tasks.

Investigating the Effect of Label Topology and Training Criterion on ASR Performance and Alignment Quality

TL;DR

The paper tackles the comparability gap between end-to-end and modular ASR by isolating the effects of label topology and training criteria. It jointly evaluates zero-order alignment models (CTC and Posterior-HMM) and first-order label-context models (Factored Hybrid HMM and Monotonic RNN-T) under fixed front-end and downsampling conditions, across LibriSpeech 960h and Switchboard 300h. Key findings show CTC can yield competitive WER and TSE on alignments, but subsequent first-order training guided by those alignments often achieves better ASR performance, with a 40 ms frame shift delivering substantial real-time factor speedups. The study demonstrates that frame-rate and topology choices materially influence both alignment quality and production-time efficiency, offering practical guidance for designing ASR pipelines that balance accuracy and latency.

Abstract

The ongoing research scenario for automatic speech recognition (ASR) envisions a clear division between end-to-end approaches and classic modular systems. Even though a high-level comparison between the two approaches in terms of their requirements and (dis)advantages is commonly addressed, a closer comparison under similar conditions is not readily available in the literature. In this work, we present a comparison focused on the label topology and training criterion. We compare two discriminative alignment models with hidden Markov model (HMM) and connectionist temporal classification topology, and two first-order label context ASR models utilizing factored HMM and strictly monotonic recurrent neural network transducer, respectively. We use different measurements for the evaluation of the alignment quality, and compare word error rate and real time factor of our best systems. Experiments are conducted on the LibriSpeech 960h and Switchboard 300h tasks.
Paper Structure (14 sections, 11 equations, 1 figure, 5 tables)

This paper contains 14 sections, 11 equations, 1 figure, 5 tables.

Table of Contents

  1. Introduction
  2. Modeling Approach
  3. Unifying Principles
  4. Mapping of the Alignment Path
  5. Conceptual Framework
  6. Models
  7. Zero-Order Label Context Alignment Models
  8. First-Order Label Context ASR Models
  9. Experimental Setting
  10. Evaluation
  11. Alignment Quality
  12. ASR Performance
  13. Conclusions
  14. Acknowledgements

Figures (1)

  • Figure 1: comparison of WER on dev-other between two P-HMMs with 10ms and 40ms frame shift when using acoustic lookahead temporal approximation scale (ALTAS). Note that the scale value represents a different temporal range for each curve.