Authors:: Felix Wu, Kwangyoun Kim, Shinji Watanabe, Kyu Han, Ryan McDonald, Kilian Q. Weinberger, Yoav Artzi
Paper link: https://arxiv.org/abs/2205.01086
| Model | Pre-training updates | Dataset | Link |
|---|---|---|---|
| Wav2Seq (from HuBERT-base) | 400K + 25K | LibriSpeech 960h | Download |
| Wav2Seq (from HuBERT-base) | 400K + 100K | LibriSpeech 960h | Download |
| Wav2Seq (from HuBERT-large) | 400K + 25K | LibriLight 60Kh | Download |
| Wav2Seq (from XLS-R 0.3B) | 400K + 25K | XLS-R multi-lingual corpus 436Kh -> LibriLight 60Kh | Download |
| Model | Pre-training updates | Finetuning split | Link |
|---|---|---|---|
| Wav2Seq (from HuBERT-base) | 400K + 25K | LibriSpeech 10h | Download |
| Wav2Seq (from HuBERT-base) | 400K + 100K | LibriSpeech 10h | Download |
| Wav2Seq (from HuBERT-base) | 400K + 25K | LibriSpeech 100h | Download |
| Wav2Seq (from HuBERT-base) | 400K + 100K | LibriSpeech 100h | Download |
| Number of Clusters | Link |
|---|---|
| 25 | Download |
| 100 | Download |
| 500 | Download |
| Number of Clusters | Number of Subwords | Link |
|---|---|---|
| 25 | 1000 | Download |
| 25 | 3000 | Download |
| 25 | 10000 | Download |
| 25 | 30000 | Download |
| 100 | 3000 | Download |
| 100 | 10000 | Download |
| 100 | 30000 | Download |
| 500 | 3000 | Download |
| 500 | 10000 | Download |
| 500 | 30000 | Download |
The code is tested with fairseq commit bba000d.
torch==1.9.0+cu111
torchaudio==0.9.0
tqdm==4.62.3
hydra-core==1.0.7
omegaconf==2.0.6
einops==0.3.0
fire==0.4.0
fairseq==1.0.0a0+bba000d
git clone [email protected]:asappresearch/wav2seq.git
cd wav2seq
pip install -e .
- Create wav2vec style manifest files
Please set
LIBRISPEECH_PATHto your librispeech folder which contains three subfolderstrain-clean-100,train-clean-360,train-other-500.
mkdir -p manifest/librispeech/train-960
python -m examples.wav2vec.wav2vec_manifest LIBRISPEECH_PATH --dest manifest/librispeech/train-960 --ext flac --valid-percent 0.01 --path-must-contain trainWe also provide our manifest here in case researcher want to use the same split. Please modify the first line of all the tsv files to ensure that the path of the data is set correctly.
- Train k-means model and get cluster indices.
Please make sure that you have download pre-trained hubert-base checkpoint at
HUBERT_PATH. Notably, this step requires a GPU for feature extraction and 64GB main memory for k-means training. Extracting HuBERT features takes about 15 minutes, training k-means may take about an hour, dumping the cluster ids of the whole Librispeech 960h data takes more than two hours.
HUBERT_PATH="save/pretrained/hubert_base_ls960.pt"
mkdir -p save/pretrained
if ! [ -f $HUBERT_PATH ]; then
wget https://dl.fbaipublicfiles.com/hubert/hubert_base_ls960.pt -O $HUBERT_PATH
fi
bash scripts/pl/extract-hubert-features.sh $HUBERT_PATH 9 2 2 500where 9, 2, 2, 500 means that we use the 9-th layer of HuBERT, kernel size 2 and stride size 2 for average pooling, and 500 custers in k-means.
- Training BPE model and create pseudo subword tokens.
bash scripts/pl/create-hubert-pseudo-language.sh labels/hubert_base-l9-k2s2-fp16-ls0.1/c500 30000We also provide our extract pseudo language here for reproducibility.
bash scripts/wav2seq-pt.sh wav2seq-hubert-base-ls960To fine-tune a pretrained checkpoint on librispeech with 10h data. Please use this command.
bash scripts/wav2seq-ft-ls.sh $pretrained_ckpt ft-ls-10hwhere $pretrained_ckpt is your pretrained checkpoint.
With 100h supervised data, please use this command.
bash scripts/wav2seq-ft-ls.sh $pretrained_ckpt ft-ls-100hPlease make sure that your manifest files are stored in manifest/librispeech.
We provide our manifest here for reproducibility. Please make sure that you change the first line of all tsv files so that the path of the data is set correctly.
We use a pretrained subword tokenizer link to convert LibriSpeech transcripts into subword tokens.