collabora/whisperspeech-librilight

This is a processed LibriLight dataset ready for training the WhisperSpeech models.

See https://github.com/collabora/WhisperSpeech for more details.

Quick start

If you want to quickly train a basic WhisperSpeech model you can start by downloading the small subset:

# magic includes to download only the small and validation data splits and the accompanying config files
huggingface-cli download --repo-type dataset --include '*-small-*' '*small.dataset' '*-speakers*' --local-dir . -- collabora/whisperspeech-librilight

# download the semantic token model to extract the token embeddings from it
huggingface-cli download collabora/whisperspeech whisper-vq-stoks-medium-en+pl.model

# the T2S training invocation:
python3 -m whisperspeech.train_multi \
  --task "t2s_up_wds_mlang_enclm base --frozen_embeddings_model whisper-vq-stoks-medium-en+pl.model" \
  --batch-size 32 --accumulate-grad-batches 2 \
  --epochs 2 --lr-schedule wsd \
  --tunables="--cps_input --causal_encoder --warmup_steps=300 --encoder_depth_ratio=.25" \
  --dataset-config=--vq_codes=513 \
  --training-data @librilight-t2s-train-small.dataset \
  --validation-data @librilight-t2s-val-common-speakers.dataset \
  --validation-data @librilight-t2s-val-unseen-speakers.dataset \
  --monitored-metric 'val_loss/dataloader_idx_0'

# the S2A training invocation:
python3 -m whisperspeech.train_multi \
  --task "s2a_delar_mup_wds_mlang tiny --quantizers 4 --spk_width=192 --frozen_embeddings_model whisper-vq-stoks-medium-en+pl.model" \
  --batch-size 48 \
  --epochs 4 --lr-schedule wsd \
  --tunables="--rope --warmup_steps=300" \
  --dataset-config=--vq_codes=513 \
  --training-data @librilight-s2a-train-small.dataset \
  --validation-data @librilight-s2a-val-common-speakers.dataset \
  --validation-data @librilight-s2a-val-unseen-speakers.dataset \
  --monitored-metric 'val_loss/dataloader_idx_0'

The --accumulate-grad-batches option is set to get a good effective batch size a single 4090 GPU. If you have multiple GPUs it will probably make sense to lower the batch size. For example 16 GPUs with a batch size of 16 seem to be give good performance and fast training.

Because we use Maximum Update Parametrization, higher effective batch sizes always result in lower losses and you don't need to adjust the learning rate. Unfortunately the effect is not linear so there is an optimal batch size and there is little benefit to increase it further.