metadata

language: en
tags:
  - Robust ASR
  - audio-to-audio
  - speech-enhancement
  - PyTorch
  - speechbrain
license: apache-2.0
datasets:
  - Voicebank
  - DEMAND
metrics:
  - WER
  - PESQ
  - eSTOI

1D CNN + Transformer Trained w/ Mimic Loss

This repository provides all the necessary tools to perform enhancement and robust ASR training (EN) within SpeechBrain. For a better experience we encourage you to learn more about SpeechBrain. The model performance is:

Release	Test PESQ	Test eSTOI	Valid WER	Test WER
21-03-08	2.92	85.2	3.20	2.96

Pipeline description

The mimic loss training system consists of three steps:

A perceptual model is pre-trained on clean speech features, the same type used for the enhancement masking system.
An enhancement model is trained with mimic loss, using the pre-trained perceptual model.
A large ASR model pre-trained on LibriSpeech is fine-tuned using the enhancement front-end.

The enhancement and ASR models can be used together or independently.

Install SpeechBrain

First of all, please install SpeechBrain with the following command:

pip install speechbrain

Please notice that we encourage you to read our tutorials and learn more about SpeechBrain.

Pretrained Usage

To use the mimic-loss-trained model for enhancement, use the following simple code:

import torchaudio
from speechbrain.pretrained import SpectralMaskEnhancement

enhance_model = SpectralMaskEnhancement.from_hparams(
    source="speechbrain/mtl-mimic-voicebank",
    savedir="pretrained_models/mtl-mimic-voicebank",
)
noisy = enhance_model.load_audio("speechbrain/mtl-mimic-voicebank/example.wav")

# Add batch dimension and enhance
enhanced = enhance_model.enhance_batch(noisy.unsqueeze(0))

# Saving enhanced signal on disk
torchaudio.save('enhanced.wav', enhanced.cpu(), 16000)

The system is trained with recordings sampled at 16kHz (single channel). The code will automatically normalize your audio (i.e., resampling + mono channel selection) when calling enhance_file if needed. Make sure your input tensor is compliant with the expected sampling rate if you use enhance_batch as in the example.

Inference on GPU

To perform inference on the GPU, add run_opts={"device":"cuda"} when calling the from_hparams method.

Training

The model was trained with SpeechBrain (150e1890). To train it from scratch follows these steps:

Clone SpeechBrain:

git clone https://github.com/speechbrain/speechbrain/

Install it:

cd speechbrain
pip install -r requirements.txt
pip install -e .

Run Training:

cd  recipes/Voicebank/MTL/ASR_enhance
python train.py hparams/enhance_mimic.yaml --data_folder=your_data_folder

You can find our training results (models, logs, etc) here.

Limitations

The SpeechBrain team does not provide any warranty on the performance achieved by this model when used on other datasets.

Referencing Mimic Loss

If you find mimic loss useful, please cite:

@inproceedings{bagchi2018spectral,
title={Spectral Feature Mapping with Mimic Loss for Robust Speech Recognition},
author={Bagchi, Deblin and Plantinga, Peter and Stiff, Adam and Fosler-Lussier, Eric},
booktitle={IEEE Conference on Audio, Speech, and Signal Processing (ICASSP)},
year={2018}
}

About SpeechBrain

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

Citing SpeechBrain

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

@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}
}