notes:
python academic/train.py -c exp_configs/experiment_soundstream_2_debug.yaml
- scripts to get LibriTTS are in data/get_data_scripts
- to run an experiment run python run_test.py -c ./exp_configs/sound_reprod_lucid.yaml -n test.
- c -config path
- n - experiment / run name
- exp_configs contains degub flag to test on dummy dataset
TODO:
- Add our loggin and wand db
- Remove model saving from trainer, move it to logger
- Add exp_name saving into model dict
- Check coding from SoundStream
- Remove installable ResidualVQ, add raw code files
- Train on a bigger dataset
- Compar results and metrics
Personal Neural Codec
This project aims to build a machine learning model for encoding and decoding voice data. The model focuses on learning efficient representations of voice while maintaining high-quality audio reconstruction.
Experiments
The following table summarizes the experiments conducted in this project:
| Experiment Name | Experiment Description |
|---|---|
| Experiment 1 | VQ-VAE model trained on mel spectrograms for audio encoding/decoding |
| Experiment 2 | TBD |
| Experiment 3 | TBD |