I’m open-sourcing a speech model training pipeline built on Burn. It trains a pronunciation scoring model from raw audio using CTC loss, with MFCC feature extraction, LibriSpeech data loading, and evaluation against human-labeled data (SpeechOcean762).

I couldn’t find any speech training examples for Burn when I needed one, so I’m putting this out as a reference. 138 tests, runs on CPU or CUDA, MIT/Apache-2.0.

Quick start:

git clone <https://github.com/dnvt/burn-speech-training>
cd burn-speech-training
cargo run --example train_small --features ndarray --release

Repo: GitHub - dnvt/burn-speech-training: End-to-end speech model training pipeline built on Burn — MFCC features, CTC loss, LibriSpeech loader, SpeechOcean762 evaluation · GitHub

Includes a 35-experiment log with what worked and what didn’t — might be useful if you’re doing anything with CTC training in Burn.

1 post - 1 participant

Read full topic