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

We offer two docker containers that are available on dockerhub and that provide the RobinASR out of the box:

• for running on GPU:

docker pull racai/robinasr:gpu
docker run --gpus all -p 8888:8888 --net=host --ipc=host racai/robinasr:gpu

• for running on CPU:

docker pull racai/robinasr:cpu
docker run -p 8888:8888 --net=host --ipc=host racai/robinasr:cpu

You can also create your own docker image by following these steps:

1. 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.

2. Build the docker image using the Dockerfile. Make sure that deepspeech_pytorch/configs/inference_config.py has the desired configuration.

docker build --tag RobinASR .

3. Run the docker image.

docker run --gpus all -p 8888:8888 --net=host --ipc=host RobinASR

1. 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.

2. 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 .

3. Install Nvidia Apex:

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

4. 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 .

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

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

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

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}
}