Views
No views yet
Self-supervised speech pre-training enables deep neural network models to capture meaningful and disentangled factors from raw waveform signals. The learned universal speech representations can then be used across numerous downstream tasks. These representations, however, are sensitive to distribution shifts caused by environmental factors, such as noise and/or room reverberation. Their large sizes, in turn, make them unfeasible for edge applications. In this work, we propose a knowledge distillation methodology termed RobustDistiller which compresses universal representations while making them more robust against environmental artifacts via a multi-task learning objective. The proposed layer-wise distillation recipe is evaluated on top of three well-established universal representations, as well as with three downstream tasks. Experimental results show the proposed methodology applied on top of the WavLM Base+ teacher model outperforming all other benchmarks across noise types and levels, as well as reverberation times. Oftentimes, the obtained results with the student model (24M parameters) achieved results inline with those of the teacher model (95M).
| Model | Parameters | #Layers | Teacher | Checkpoint |
|---|---|---|---|---|
| RD (WavLM) | 27M | 2 | WavLM | link |
| RD (HuBERT) | 27M | 2 | HuBERT | link |
| RD (Robust HuBERT) | 27M | 2 | Robust HuBERT | link |
| RD (DPWavLM) | 27M | 12 | WavLM | link |
pip install -U transformers1import torch
2import torchaudio
3from transformers import AutoModel
4
5# Load pre-trained model
6model = AutoModel.from_pretrained("Hguimaraes/rd_wavlm", trust_remote_code=True).cuda().eval()
7
8# Load audio and resample to 16kHz
9wav, sr = torchaudio.load_audio("path/to/audio") # (batch_size, wav_len)
10wav = torchaudio.functional.resample(wav, sr, 16000)
11
12# Extract features
13with torch.no_grad():
14 output = model(wav)
15
16# output["last_hidden_states"]: final output (batch_size, seq_len, encoder_dim)
17# output["hidden_states"]: list of elements with (batch_size, seq_len, encoder_dim) tensors (features for each layer or projections)1@inproceedings{guimaraes2023robustdistiller,
2 title={RobustDistiller: Compressing Universal Speech Representations for Enhanced Environment Robustness},
3 author={Guimarães, Heitor R and Pimentel, Arthur and Avila, Anderson R and Rezagholizadeh, Mehdi and Chen, Boxing and Falk, Tiago H},
4 booktitle={ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
5 pages={1--5},
6 year={2023},
7 organization={IEEE}
8}
9
10@article{guimaraes2024efficient,
11 title={An Efficient End-to-End Approach to Noise Invariant Speech Features via Multi-Task Learning},
12 author={Guimar{\~a}es, Heitor R and Pimentel, Arthur and Avila, Anderson R and Rezagholizadeh, Mehdi and Chen, Boxing and Falk, Tiago H},
13 journal={arXiv preprint arXiv:2403.08654},
14 year={2024}
15}