SEW (Squeezed and Efficient Wav2vec)

The repo contains the code of the paper "Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition" by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q Weinberger, and Yoav Artzi.

Model Checkpoints

Unsupervisedly Pre-trained on LibriSpeech 960h

ASR model fine-tuned on LibriSpeech train-clean 100h

Usage

Dependencies

The code is tested with fairseq commit 05255f9, deberta commit bf17ca4 and the following packages.

torch==1.8.0
torchaudio==0.8.0
tqdm==4.49.0
Hydra==2.5
hydra-core==1.0.4
fvcore==0.1.5.post20210330
omegaconf==2.0.5
einops==0.3.0
fire==0.2.1

Apex

Please install NVIDIA's apex with

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" \
  --global-option="--deprecated_fused_adam" --global-option="--xentropy" \
  --global-option="--fast_multihead_attn" ./

wav2letter decoder

Currently, we are decoding with wav2letter v0.2 python binding at commit 96f5f9d Please install the python binding here https://github.com/flashlight/wav2letter/tree/96f5f9d3b41e01af0a031ee0d2604acd9ef3b1b0/bindings/python The newest commit d5a93f0 in v0.2 branch leads to worse WER for wav2vec 2.0 baselines.

Installation

git clone https://github.com/asappresearch/sew.git
cd sew 
pip install -e .

Pre-training

Pre-training SEW models

Run the following command where $model_size can be tiny, small, or mid, and $ngpu is tne number of GPUs you want to use.

bash scripts/pt-sew.sh $model_size $ngpu

Pre-training SEW-D models

bash scripts/pt-sew-d.sh $model_size $ngpu

where $model_size can be tiny, small, mid, mid-k127, base, or base+.

Fine-tuning

Run the following script to fine-tune a model with the hyperparameters from wav2vec 2.0.

bash scripts/ft-model.sh $pre_trained_model $split $ngpu

where $pre_trained_model can be either a W2V2, SEW, or a SEW-D model checkpoint and $split can be 10m, 1h, 10h, or 100h.

Here we also provide a set of hyperparameters which sets all dropouts the same as the pre-training stage, and we found it to be more stable.

bash scripts/ft-model-stable.sh $pre_trained_model $split $ngpu

If you see out of GPU memory error, please scale down the dataset.max_tokens and scale up the optimization.update_freq in scripts/ft-model.sh. For example modifying these lines

  dataset.max_tokens=3200000 \
  optimization.update_freq="[$((8 / $ngpu))]" \

to

  dataset.max_tokens=1600000 \
  optimization.update_freq="[$((16 / $ngpu))]" \

which reduces the batch size and increases the gradient accumulation steps in order to use less GPU memory.

Evaluation

1. Please run this script to prepare the official LibriSpeech 4-gram language model.

bash scripts/prepare_librispeech_lm.sh $kenlm_build_bin

where $kenlm_build_bin is the folder that contains the KenLM build_binary executable file (e.g. /home/user/kenlm/build/bin).

2. Then run this script to evaluate a pre-trained ASR model

python tools/eval_w2v.py tunelm --subsets '["dev-clean", "dev-other", "test-clean", "test-other"]' --model $asr_checkpoint