This recipe combines enhancement and ASR to improve performance on both tasks. The technique we use in this recipe is a perceptual loss with a speech recognizer, which we have called mimic loss [1, 2, 3] and is performed in three main stages:

1. Pretrain an acoustic model as a perceptual model of speech, used to judge the perceptual quality of the outputs of the enhancement model.
2. Train an enhancement model by freezing the perceptual model, passing clean and enhanced features to the perceptual model, and generating a loss using the MSE between the outputs of the perceptual model.
3. Freezing the enhancement model and training a robust ASR model to recognize the enhanced outputs.

This approach is similar to joint training of enhancement and ASR models, but maintains the advantages of interpretability and independence, since each model can be used for other data or tasks without requiring the co-trained model.

Before proceeding, ensure you have installed the necessary additional dependencies. To do this, simply run the following command in your terminal:

pip install -r extra_requirements.txt

To train these models from scratch, you can run these three steps using the following commands:

> python train.py hparams/pretrain_perceptual.yaml
> python train.py hparams/enhance_mimic.yaml
> python train.py hparams/robust_asr.yaml

One important note is that each step depends on one or more pretrained models, so ensuring these exist and the paths are correct is an important step. The path in hparams/enhance_mimic.yaml should point at the src_embedding.ckpt model trained in step 1, and the path in hparams/enhance_mimic.yaml should point at the enhance_model.ckpt model trained in step 2.

Joint training can be achieved by adding the enhance_model to the "unfrozen" models so that the weights are allowed to update. To see enhancement scores, add an enhancement loss after training is complete and run the script again.

The PESQ and eSTOI results are generated using the test set, and the WER results are generated over 3 runs. The last 5 epochs are combined so no validation data is used to choose checkpoints.

Results generated using updated Wide ResNet from [2, 3]. Additions include:

1. Squeeze-and-excitation blocks
2. Spectral approximation algorithm on complex spectrogram
3. 2d batch normalization
4. GELU activations

You can find the pre-trained model with an easy-inference function on HuggingFace: https://huggingface.co/speechbrain/mtl-mimic-voicebank

You can find the full experiment folder (i.e., checkpoints, logs, etc) here: https://www.dropbox.com/sh/azvcbvu8g5hpgm1/AACDc6QxtNMGZ3IoZLrDiU0Va?dl=0

[1] Deblin Bagchi, Peter Plantinga, Adam Stiff, Eric Fosler-Lussier, “Spectral Feature Mapping with Mimic Loss for Robust Speech Recognition.” ICASSP 2018 https://arxiv.org/abs/1803.09816

[2] Peter Plantinga, Deblin Bagchi, Eric Fosler-Lussier, “An Exploration of Mimic Architectures for Residual Network Based Spectral Mapping.” SLT 2018 https://arxiv.org/abs/1809.09756

[3] Peter Plantinga, Deblin Bagchi, Eric Fosler-Lussier, “Phonetic Feedback For Speech Enhancement With and Without Parallel Speech Data.” ICASSP 2020 https://arxiv.org/abs/2003.01769

• Website: https://speechbrain.github.io/
• Code: https://github.com/speechbrain/speechbrain/
• HuggingFace: https://huggingface.co/speechbrain/

Please, cite SpeechBrain if you use it for your research or business.

@misc{speechbrainV1,
  title={Open-Source Conversational AI with SpeechBrain 1.0},
  author={Mirco Ravanelli and Titouan Parcollet and Adel Moumen and Sylvain de Langen and Cem Subakan and Peter Plantinga and Yingzhi Wang and Pooneh Mousavi and Luca Della Libera and Artem Ploujnikov and Francesco Paissan and Davide Borra and Salah Zaiem and Zeyu Zhao and Shucong Zhang and Georgios Karakasidis and Sung-Lin Yeh and Pierre Champion and Aku Rouhe and Rudolf Braun and Florian Mai and Juan Zuluaga-Gomez and Seyed Mahed Mousavi and Andreas Nautsch and Xuechen Liu and Sangeet Sagar and Jarod Duret and Salima Mdhaffar and Gaelle Laperriere and Mickael Rouvier and Renato De Mori and Yannick Esteve},
  year={2024},
  eprint={2407.00463},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2407.00463},
}
@misc{speechbrain,
  title={{SpeechBrain}: A General-Purpose Speech Toolkit},
  author={Mirco Ravanelli and Titouan Parcollet and Peter Plantinga and Aku Rouhe and Samuele Cornell and Loren Lugosch and Cem Subakan and Nauman Dawalatabad and Abdelwahab Heba and Jianyuan Zhong and Ju-Chieh Chou and Sung-Lin Yeh and Szu-Wei Fu and Chien-Feng Liao and Elena Rastorgueva and François Grondin and William Aris and Hwidong Na and Yan Gao and Renato De Mori and Yoshua Bengio},
  year={2021},
  eprint={2106.04624},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  note={arXiv:2106.04624}
}