aaaceo890 / abc_asr Goto Github PK

This respository is the official implementation of "End-to-End Multi-Modal Speech Recognition on an Air and Bone Conducted Speech Corpus" for TASLP 2023.

1. If you just need the module only, run

   pip install espnet
   

   first, and you can use the modules in abc_asr/model.

2. If you want to do full experiments, you need to correctly install ESPnet and kaldi first. See Installation.

   Next, run

   pip install -r requirements.txt
   

   to install the required packages.

1. Download dataset.

   Download the ABCS Corpus here: Links.

   Download the noisy air conducted data (ns_air_data.zip) here: [Onedrive] or [Baidu Cloud]

   Unzip the noisy data into ABCS's directory:

   unzip -d <ABCS dir>/Audio/ ns_air_data.zip
   

2. Execute the data preparation script.

   For inference only:

   python3 data_prep --dataset_root <ABCS dir> --test
   

   For full experiments:

   python3 data_prep --dataset_root <ABCS dir>
   

1. Ensure that kaldi and ESPnet are properly installed on your environment. Next, have correctly adjust the third line in test.sh:

   export ESPNETROOT=<Your Espnet Root>
   

2. Download the model parameters file here [Onedrive] or [Baidu Cloud]

   mv model.acc.best <Your Path>/abc_asr/results
   

3. Run

   bash test.sh
   

The training pipeline.

If you found this code helpful, please consider citing it as follows:

@ARTICLE{9961873,
  author={Wang, Mou and Chen, Junqi and Zhang, Xiao-Lei and Rahardja, Susanto},
  journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing}, 
  title={End-to-End Multi-Modal Speech Recognition on an Air and Bone Conducted Speech Corpus}, 
  year={2023},
  volume={31},
  number={},
  pages={513-524},
  keywords={Speech recognition;Speech processing;Signal to noise ratio;Spectrogram;Headphones;Microphones;Synchronization;Speech recognition;multi-modal speech processing;bone conduction;air- and bone-conducted speech corpus},
  doi={10.1109/TASLP.2022.3224305}}