Phone Classification using Wav2Vec2

This repository contains Speechbrain recipes to fine-tune Wav2Vec2 models on a phone classification task. Following factors were analysed:

1. Fine-tuning Wav2Vec2,
2. Pre-training datasets,
3. Model size,
4. fine-tuning datasets.

Results of this work have been published at the Interspeech 2024 conference.

Code

• The recipes folder contains all Speechbrain recipes.
• Results obtained are available in the confusion-matrix/ folder.

Data

For confidentiality reasons, datasets are not included. This work relies on the C2SI, CommonPhone and BREF corpora.

Recipes

Details of some of the Speechbrain recipes set up in this repository.

• unfrozen-cp-3k-large-accents is the best recipe published in the Interspeech paper listed below.
• unfrozen-cp-3k-large-accents-argmax takes the maximum of all 6 segments (1024-dim). LeakyReLu.
• unfrozen-cp-3k-large-concatenate take both central segments (2048-dim) as input to the classifier.

How to cite

If you use this work, please cite as:

@inproceedings{maisonneuve24,
  title     = {Towards objective and interpretable speech disorder assessment: a comparative analysis of CNN and transformer-based models},
  author    = {Malo Maisonneuve and Corinne Fredouille and Muriel Lalain and Alain Ghio and Virginie Woisard},
  year      = {2024},
  booktitle = {Interspeech 2024},
  pages     = {1970--1974},
  doi       = {10.21437/Interspeech.2024-267},
  issn      = {2958-1796},
}