Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Back to Supervision: Boosting Word Boundary Detection through Frame Classification

Simone Carnemolla, Salvatore Calcagno, Simone Palazzo, Daniela Giordano

TL;DR

A model-agnostic framework to perform word boundary detection in a supervised manner also employing a labels augmentation technique and an output-frame selection strategy, which surpasses the performance of other state-of-the-art architectures whether trained in supervised or self-supervised settings on the same datasets.

Abstract

Speech segmentation at both word and phoneme levels is crucial for various speech processing tasks. It significantly aids in extracting meaningful units from an utterance, thus enabling the generation of discrete elements. In this work we propose a model-agnostic framework to perform word boundary detection in a supervised manner also employing a labels augmentation technique and an output-frame selection strategy. We trained and tested on the Buckeye dataset and only tested on TIMIT one, using state-of-the-art encoder models, including pre-trained solutions (Wav2Vec 2.0 and HuBERT), as well as convolutional and convolutional recurrent networks. Our method, with the HuBERT encoder, surpasses the performance of other state-of-the-art architectures, whether trained in supervised or self-supervised settings on the same datasets. Specifically, we achieved F-values of 0.8427 on the Buckeye dataset and 0.7436 on the TIMIT dataset, along with R-values of 0.8489 and 0.7807, respectively. These results establish a new state-of-the-art for both datasets. Beyond the immediate task, our approach offers a robust and efficient preprocessing method for future research in audio tokenization.

Back to Supervision: Boosting Word Boundary Detection through Frame Classification

TL;DR

A model-agnostic framework to perform word boundary detection in a supervised manner also employing a labels augmentation technique and an output-frame selection strategy, which surpasses the performance of other state-of-the-art architectures whether trained in supervised or self-supervised settings on the same datasets.

Abstract

Speech segmentation at both word and phoneme levels is crucial for various speech processing tasks. It significantly aids in extracting meaningful units from an utterance, thus enabling the generation of discrete elements. In this work we propose a model-agnostic framework to perform word boundary detection in a supervised manner also employing a labels augmentation technique and an output-frame selection strategy. We trained and tested on the Buckeye dataset and only tested on TIMIT one, using state-of-the-art encoder models, including pre-trained solutions (Wav2Vec 2.0 and HuBERT), as well as convolutional and convolutional recurrent networks. Our method, with the HuBERT encoder, surpasses the performance of other state-of-the-art architectures, whether trained in supervised or self-supervised settings on the same datasets. Specifically, we achieved F-values of 0.8427 on the Buckeye dataset and 0.7436 on the TIMIT dataset, along with R-values of 0.8489 and 0.7807, respectively. These results establish a new state-of-the-art for both datasets. Beyond the immediate task, our approach offers a robust and efficient preprocessing method for future research in audio tokenization.

Paper Structure

This paper contains 22 sections, 10 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Method
  4. Overview
  5. Problem Formulation
  6. Labels Processing
  7. Model Architecture
  8. Post-processing
  9. Experimental setup
  10. Datasets
  11. Audio Pre-processing
  12. Encoders
  13. Training Procedure
  14. Metrics
  15. Experiments
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Overview of our method
  • Figure 2: Segmentation comparison example between true (green solid line) and predicted (red dashed line) boundaries on Buckeye test set. The four pictures on top show a correct match of detected boundaries. The second group report three under segmentation scenarios and an over segmentation one.
  • Figure 3: Comparison of R-value and F-value for the Buckeye validation set based on different window sizes for label augmentation. Each data point represents the number of frames labeled as begin to the left and right of the ground truth. The results are computed employing the HuBERT encoder.
  • Figure 4: Comparison of R-value and F-value scores for the Buckeye validation set based on different frame selection strategies. The first approach retrieves the initial begin frame from the begin cluster, the second approach selects the middle frame, and the third approach picks the final frame. The results are computed employing the HuBERT encoder.