Awesome

Losses Can Be Blessings: Routing Self-Supervised Speech Representations Towards Efficient Multilingual and Multitask Speech Processing

Yonggan Fu, Yang Zhang, Kaizhi Qian, Zhifan Ye, Zhongzhi Yu, Cheng-I Lai, Yingyan (Celine) Lin

Accepted at NeurIPS 2022. [ Paper | Video | Slide ]

S³-Router: Overview

• Self-supervised learning (SSL) for speech representations has achieved empirical success in low-resource Automatic Speech Recognition, while the resultant speech SSL models have become prohibitively large, limiting their deployment on resource-constrained devices. Additionally, strongly overparameterized speech SSL models tend to suffer from overfitting. This work aims to enhance the practical usage of speech SSL models towards a win-win in both enhanced efficiency and alleviated overfitting via our proposed S³-Router framework, which can serve as an all-in-one technique to enable (1) a new finetuning paradigm, (2) an efficient multilingual/multitask solution, (3) a SOTA ASR pruning scheme, and (4) a new tool to analyze the learned speech representation.

S³-Router: Framework

• Key insights: Model sparsity can not only favor efficiency, but also can serve as a similar role, i.e., another optimization knob, as model weights to encode language-/task-specific information.

• Key idea: Finetune model connections instead of model weights via optimizing language-/task-specific binary masks on top of shared weights inherited from pretrained speech SSL models.

S³-Router's Application 1: A New Finetuning Paradigm

• Discarding ≤10% weights is all you need: Consistently outperform the standard weight finetuning in the achievable word error rate.

• Insight: Tuning model connections instead of weights can reduce overfitting on low-resource speech.

S³-Router's Application 2: An Efficient Multilingual/Multitask Solution

• Simultaneously support the aforementioned 11 languages (10 languages below + English) with -88.5% parameters.

S³-Router's Application 3: A SOTA Pruning Scheme

• Achieve better or comparable pruning effectiveness over SOTA ASR pruning techniques across different resource settings.

S³-Router's Application 4: Analyze Speech SSL Models

• The similarity among binary masks learned on different languages by speech SSL models is highly correlated with their phonetic similarity defined by human experts.

Code Usage

Our code is built on top of [Fairseq].

Installation

• To install fairseq and develop locally:

pip install --editable ./

• For faster training install NVIDIA's apex library:

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" ./

• Prepare the LibriSpeech/CommonVoice datasets following wav2vec 2.0 in Fairseq.

Finetune wav2vec 2.0 @ LibriSpeech via S³-Router

• Finetune wav2vec 2.0 on LibriSpeech-1h with a sparsity of 10%:

CUDA_VISIBLE_DEVICES=0,1 fairseq-hydra-train task.data=path-to-librispeech  model.w2v_path=path-to-wav2vec_small.pt dataset.train_subset=train-1h dataset.valid_subset=test-clean hydra.run.dir=outputs/wav2vec2-base dataset.max_tokens_valid=800000  dataset.max_tokens=1200000 distributed_training.distributed_init_method=tcp://localhost:15460 model._name=wav2vec_ctc_st model.prune_rate=0.9 checkpoint.no_epoch_checkpoints=false checkpoint.save_interval=1000 optimization.lr=[0.00005] model.fix_attn=true model.trainable_proj=true lr_scheduler.final_lr_scale=0.0001 model.init_score=weight_rank --config-dir examples/wav2vec/config/finetuning --config-name base_1h

Finetune data2vec @ LibriSpeech via S³-Router

• Finetune data2vec on LibriSpeech-1h with a sparsity of 10%:

CUDA_VISIBLE_DEVICES=0,1 fairseq-hydra-train task.data=path-to-librispeech model.w2v_path=path-to-audio_base_ls.pt dataset.train_subset=train-1h dataset.valid_subset=test-clean hydra.run.dir=outputs/data2vec dataset.max_tokens_valid=800000  dataset.max_tokens=1200000 distributed_training.distributed_init_method=tcp://localhost:15460 task.normalize=true model._name=wav2vec_ctc_st model.prune_rate=0.9 checkpoint.no_epoch_checkpoints=false checkpoint.save_interval=1000 optimization.lr=[0.00005] model.fix_attn=true model.trainable_proj=true lr_scheduler.final_lr_scale=0.0001 common.user_dir=examples/data2vec --config-dir examples/wav2vec/config/finetuning --config-name base_1h

Finetune xlsr @ CommonVoice via S³-Router

• Finetune xlsr on Mandarin @ CommonVoice with a sparsity of 6%:

CUDA_VISIBLE_DEVICES=0,1 fairseq-hydra-train task.data=path-to-commonvoice-zh_TW-train model.w2v_path=path-to-xlsr2_300m.pt dataset.train_subset=zh_TW-train dataset.valid_subset=zh_TW-test hydra.run.dir=outputs/xlsr dataset.max_tokens_valid=800000 dataset.max_tokens=1200000 distributed_training.distributed_init_method=tcp://localhost:15460 model._name=wav2vec_ctc_st model.prune_rate=0.94 model.checkpoint_activations=true checkpoint.no_epoch_checkpoints=false checkpoint.save_interval=1000 optimization.lr=[0.00005] model.trainable_proj=true model.fix_attn=true checkpoint.best_checkpoint_metric=uer  model.init_score=weight_rank lr_scheduler.final_lr_scale=0.0001 --config-dir examples/wav2vec/xlsr/config --config-name finetune

Pruning wav2vec 2.0 @ LibriSpeech via S³-Router

• [Step 1] Finetune wav2vec 2.0 using standard weight finetuning on LibriSpeech-1h:

CUDA_VISIBLE_DEVICES=0,1 fairseq-hydra-train task.data=path-to-librispeech  model.w2v_path=path-to-wav2vec_small.pt dataset.train_subset=train-1h     dataset.valid_subset=test-clean hydra.run.dir=outputs/finetune dataset.max_tokens_valid=800000  dataset.max_tokens=1200000 distributed_training.distributed_init_method=tcp://localhost:15460 model._name=wav2vec_ctc checkpoint.no_epoch_checkpoints=false checkpoint.save_interval=1000 optimization.lr=[0.00005] --config-dir examples/wav2vec/config/finetuning --config-name base_1h

• [Step 2] Prune finetuned wav2vec 2.0 on LibriSpeech-1h via S³-Router with a sparsity of 70%:

CUDA_VISIBLE_DEVICES=0,1 fairseq-hydra-train task.data=path-to-librispeech  model.w2v_path=path-to-wav2vec_small.pt dataset.train_subset=train-1h     dataset.valid_subset=test-clean hydra.run.dir=outputs/pruning_wav2vec2 dataset.max_tokens_valid=800000  dataset.max_tokens=1200000 distributed_training.distributed_init_method=tcp://localhost:15460 model._name=wav2vec_ctc_st model.prune_rate=0.3 checkpoint.no_epoch_checkpoints=false checkpoint.save_interval=1000 optimization.lr=[0.00005] model.fix_attn=false model.trainable_proj=true lr_scheduler.final_lr_scale=0.0001 model.init_score=weight_magnitude_with_scale checkpoint.finetune_from_model=outputs/finetune/checkpoints/checkpoint_best.pt --config-dir examples/wav2vec/config/finetuning --config-name base_1h

Citation

• If you find our work interesting or helpful to your research, welcome to cite our paper:

@article{fu2022losses,
  title={Losses Can Be Blessings: Routing Self-Supervised Speech Representations Towards Efficient Multilingual and Multitask Speech Processing},
  author={Fu, Yonggan and Zhang, Yang and Qian, Kaizhi and Ye, Zhifan and Yu, Zhongzhi and Lai, Cheng-I and Lin, Yingyan},
  journal={arXiv preprint arXiv:2211.01522},
  year={2022}
}