---
language:
- de
thumbnail: null
pipeline_tag: automatic-speech-recognition
tags:
- whisper
- pytorch
- speechbrain
- Transformer
license: apache-2.0
datasets:
- RescueSpeech
metrics:
- wer
- sisnri
- sdri
- pesq
- stoi
model-index:
- name: noisy-whisper-resucespeech
  results:
  - task:
      name: Noise Robust Automatic Speech Recognition
      type: noise-robust-automatic-speech-recognition
    metrics:
      - name: Test WER
        type: wer
        value: '24.20'
      - name: Test PESQ
        type: pesq
        value: '2.085'
      - name: Test SI-SNRi
        type: si-snri
        value: '7.334'
      - name: Test SI-SDRi
        type: si-sdri
        value: '7.871'
inference: false
---

# Noise robust speech recognition on jointly trained SepFormer speech enhancement and Whisper ASR using RescueSpeech data.

This repository provides all the necessary tools to perform noise robust automatic speech
recognition on a simple combination of an enhancement model (**SepFormer**) and speech recognizer (**Whisper**). 
Initially, the models are fine-tuned individually on the RescueSpeech dataset, and then they are integrated to undergo joint training, enabling them to effectively handle noise interference. For a better experience, we encourage you to learn more about
[SpeechBrain](https://speechbrain.github.io).

The performance of the model is the following on test set:

| Release | SISNRi | SDRi | PESQ | STOI | WER | GPUs |
|:-------------:|:--------------:|:--------------:| :--------:|:--------------:| :--------:|:--------:|
| 07-11-23 | 7.482	 |  8.011	 | 2.083	| 0.859 | 45.29 | 1xA100 80 GB |

## Pipeline description
- The enhancement system is composed of SepFormer model.
  - The model is first trained on Microsoft-DNS dataset and subsequently fine-tuned on RescueSpeech dataset.
  - The enhanced utterances are fed to the ASR model.
- And the ASR system is composed of whisper encoder-decoder blocks:
  - The pretrained whisper-large-v2 encoder is frozen.
  - The pretrained Whisper tokenizer is used.
  - A pretrained Whisper-large-v2 decoder ([openai/whisper-large-v2](https://huggingface.co/openai/whisper-large-v2)) is finetuned on RescueSpeech dataset.
  The obtained final acoustic representation is given to the greedy decoder. 

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 *transcribe_file* if needed.

## Install SpeechBrain

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

```
pip install speechbrain
```

Please notice that we encourage you to read our tutorials and learn more about
[SpeechBrain](https://speechbrain.github.io).

### Transcribing your own audio files (in German)

```python
from speechbrain.inference.separation import SepformerSeparation as Separator
from speechbrain.inference.ASR import WhisperASR
import torch

enh_model = Separator.from_hparams(
                                source="speechbrain/noisy-whisper-resucespeech", 
                                savedir='pretrained_models/noisy-whisper-rescuespeech',
                                hparams_file="enhance.yaml"
                                )
asr_model = WhisperASR.from_hparams(
                                  source="speechbrain/noisy-whisper-resucespeech", 
                                  savedir="pretrained_models/noisy-whisper-rescuespeech",
                                  hparams_file="asr.yaml"
                                  )

# For custom file, change the path accordingly
est_sources = enh_model.separate_file(path='speechbrain/noisy-whisper-resucespeech/example_rescuespeech16k.wav')
pred_words, _ = asr_model(est_sources[:, :, 0], torch.tensor([1.0]))
```
### Inference on GPU
To perform inference on the GPU, add  `run_opts={"device":"cuda"}`  when calling the `from_hparams` method.


You can find our training results (models, logs, etc) [here](https://www.dropbox.com/sh/kqs2ld14fm20cxl/AACiobSLdNtXhm-4Y3IIbTeia?dl=0).

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

#### Referencing SpeechBrain

```
@misc{SB2021,
    author = {Ravanelli, Mirco and Parcollet, Titouan and Rouhe, Aku and Plantinga, Peter and Rastorgueva, Elena and Lugosch, Loren and Dawalatabad, Nauman and Ju-Chieh, Chou and Heba, Abdel and Grondin, Francois and Aris, William and Liao, Chien-Feng and Cornell, Samuele and Yeh, Sung-Lin and Na, Hwidong and Gao, Yan and Fu, Szu-Wei and Subakan, Cem and De Mori, Renato and Bengio, Yoshua },
    title = {SpeechBrain},
    year = {2021},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\\\\url{https://github.com/speechbrain/speechbrain}},
  }
```

### Referencing RescueSpeech
```bibtex
@misc{sagar2023rescuespeech,
    title={RescueSpeech: A German Corpus for Speech Recognition in Search and Rescue Domain},
    author={Sangeet Sagar and Mirco Ravanelli and Bernd Kiefer and Ivana Kruijff Korbayova and Josef van Genabith},
    year={2023},
    eprint={2306.04054},
    archivePrefix={arXiv},
    primaryClass={eess.AS}
}
```

#### About SpeechBrain
SpeechBrain is an open-source and all-in-one speech toolkit. It is designed to be simple, extremely flexible, and user-friendly. Competitive or state-of-the-art performance is obtained in various domains.

Website: https://speechbrain.github.io/

GitHub: https://github.com/speechbrain/speechbrain