Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Simple HMM with Self-Supervised Representations for Phone Segmentation

Gene-Ping Yang, Hao Tang

TL;DR

It is shown that peak detection on Mel spectrograms is a strong baseline, better than many self-supervised approaches, and a simple hidden Markov model is proposed that uses self-supervised representations and features at the boundaries for phone segmentation.

Abstract

Despite the recent advance in self-supervised representations, unsupervised phonetic segmentation remains challenging. Most approaches focus on improving phonetic representations with self-supervised learning, with the hope that the improvement can transfer to phonetic segmentation. In this paper, contrary to recent approaches, we show that peak detection on Mel spectrograms is a strong baseline, better than many self-supervised approaches. Based on this finding, we propose a simple hidden Markov model that uses self-supervised representations and features at the boundaries for phone segmentation. Our results demonstrate consistent improvements over previous approaches, with a generalized formulation allowing versatile design adaptations.

A Simple HMM with Self-Supervised Representations for Phone Segmentation

TL;DR

It is shown that peak detection on Mel spectrograms is a strong baseline, better than many self-supervised approaches, and a simple hidden Markov model is proposed that uses self-supervised representations and features at the boundaries for phone segmentation.

Abstract

Despite the recent advance in self-supervised representations, unsupervised phonetic segmentation remains challenging. Most approaches focus on improving phonetic representations with self-supervised learning, with the hope that the improvement can transfer to phonetic segmentation. In this paper, contrary to recent approaches, we show that peak detection on Mel spectrograms is a strong baseline, better than many self-supervised approaches. Based on this finding, we propose a simple hidden Markov model that uses self-supervised representations and features at the boundaries for phone segmentation. Our results demonstrate consistent improvements over previous approaches, with a generalized formulation allowing versatile design adaptations.
Paper Structure (12 sections, 7 equations, 2 figures, 4 tables)

This paper contains 12 sections, 7 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Boundary Features in Mel Spectrogram
  3. Applying HMMs to unsupervised phone segmentation
  4. HMM Formulation
  5. Boundary Features as Transition Penalty
  6. Related Work
  7. Experiments
  8. Self-supervised Features using Peak Detection
  9. Proposed HMMs
  10. HMM Training vs. Two-stage Decoding
  11. HMM Phone Purity Analysis
  12. Conclusion

Figures (2)

  • Figure 1: Peak detection using Mel spectrogram on the sample utterance fadg0_sx289 from TIMIT. From top to bottom: Mel spectrogram, spectral variations, and ground truth phone segments.
  • Figure 2: Comparison of the detected boundaries by different HMMs using HuBERT features on fadg0_si1909 from TIMIT.