RobinASR

This repository contains Robin's Automatic Speech Recognition (RobinASR) for the Romanian language based on the DeepSpeech2 architecture, together with a KenLM language model to imporve the transcriptions.

The pretrained text-to-speech model can be downloaded from here and the pretrained KenLM can be downloaded from here.

Also, make sure to visit:

A demo of the ASR system available in the RELATE platform: https://relate.racai.ro/index.php?path=robin/asr
A post-processing web service allowing hyphenation and basic capitalization restoration: https://github.com/racai-ai/RobinASRHyphenationCorrection

Installation

Download the pretrained text-to-speech model and the pretrained KenLM at the above links, and copy them in a models directory inside this repository.

From Source

You must have Python 3.6+ and PyTorch 1.5.1+ installed in your system. Also. Cuda 10.1+ is required if you want to use the (recommended) GPU version.
Clone the repository and install its dependencies:

git clone https://github.com/racai-ai/RobinASR.git
cd RobinASR
pip3 install -r requirements.txt
pip3 install -e .

Install Nvidia Apex:

git clone --recursive https://github.com/NVIDIA/apex.git
cd apex && pip install .

If you want to use Beam Search and the KenLM language model, you must install CTCDecode:

git clone --recursive https://github.com/parlance/ctcdecode.git
cd ctcdecode && pip install .

Inference Server

Firstly, take a look at the configuration file in deepspeech_pytorch/configs/inference_config.py and make sure that the configuration meets your requirements. Then, run the following command:

python3 server.py

Train a New Model

You must create 3 csv manifest files (train, valid and test) that contain on each line the the path to a wav file and the path to its corresponding transcription, separated by commas:

path_to_wav1,path_to_txt1
path_to_wav2,path_to_txt2
path_to_wav3,path_to_txt3
...

Then you must modify correspondingly with your configuration the file located at deepspeech_pytorch/configs/train_config.py and start training with:

python train.py

Acknowledgments

We would like to thank Sean Narnen for making his DeepSpeech2 implementation publicly-available. We used a lot of his code in our implementation.

Cite

If you are using this repository, please cite the following paper as a thank you to the authors:

Avram, A.M., Păiș, V. and Tufis, D., 2020, October. Towards a Romanian end-to-end automatic speech recognition based on Deepspeech2. In Proc. Rom. Acad. Ser. A (Vol. 21, pp. 395-402).

or in BibTeX format:

@inproceedings{avram2020towards,
  title={Towards a Romanian end-to-end automatic speech recognition based on Deepspeech2},
  author={Avram, Andrei-Marius and Păiș, Vasile and Tufiș, Dan},
  booktitle={Proceedings of the Romanian Academy, Series A},
  pages={395--402},
  year={2020}
}

adrianstanea/RobinASR