Tiny Voice

By Tiny Workshop
An open-source, user friendly library to fine-tune speech-to-text models on edge device CPUs.

Installation

We set up a virtual environment using conda, and our code is developed in Python 3.12.
NOTE: Tiny Voice is has only been tested on Linux systems, if using Windows, please install Windows System Linux (WSL):

# Cloning the repository
git clone https://github.com/derpysquid10/tiny_voice.git

cd tiny_voice

# Setting up environment and installing libraries
conda env create -f environment.yml
conda activate tiny-voice
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu  # We want the CPU version of pytorch
pip install -r requirements.txt

Quickstart Example

This section provides an example of using Tiny Voice to fine-tune OpenAI's Whisper Base model on the Afrispeech-200 dataset.

We first have to load the dataset:

python tiny_voice/data_processing.py --process-data

The example.py script fine-tunes the model using one of partial fine-tuning, LoRA, or IA3 on one of 2 configs of the dataset

python tiny_voice/example.py

If an error version `GLIBCXX_3.4.30' not found is encountered, run conda install -c conda-forge libstdcxx-ng=12 and re-run the baseline tests.