IST-ASR

Automatic Speech Recognition (ASR) project for the Inclusive Speech Technology course.

Project Overview

This project implements an end-to-end Automatic Speech Recognition system, featuring model training, fine-tuning, and evaluation components. The system is designed to transcribe speech audio into text with a focus on minimizing Word Error Rate (WER).

Project Structure

• experiments/: Scripts for running and analyzing ASR experiments. Used for training the baseline model

  • run_experiment.sh: Main script for launching experiments
  • test_experiment.sh: Script for testing experiment results
  • train_model.py: Python script for model training
  • batch_scripts/: Contains multiple experiment batch scripts
  • hparams/: Hyperparameter configuration files
  • logs/: Experiment log files

• fine-tuning/: Components for fine-tuning pre-trained models

  • fine-tune.py: Fine-tuning implementation
  • fine-tune.sh: Shell script wrapper for fine-tuning
  • fine-tune.yaml: Configuration for fine-tuning parameters

• random_split/: Dataset splits for training and evaluation

  • split_stats.txt: Statistics about the dataset splits
  • train.csv: Training dataset
  • val.csv: Validation dataset
  • test.csv: Test dataset

• results/: Evaluation results and metrics

  • aggregated-wer-results.txt: Compiled Word Error Rate results
  • Results organized by model variant (baseline/, fine-tuning/, model_testing/)

• testing/: Model evaluation scripts

  • test_models.py: Python script for model evaluation
  • test_models.sh: Shell script wrapper for testing

• tokenizer/: Tools for tokenization of text data

  • tokenizer.yaml: Configuration for the tokenizer
  • train_tokenizer.py: Script to train custom tokenizers
  • train_tokenizer.sh: Shell wrapper for tokenizer training