README.md

ASR

Several ASR implementations. A full description of the project can be found in the paper.

installaion

• Download the AN4 dataset.
• run pip install -r requirments.txt
• for conda: run conda env create -f environment.yml
• experiment with the models!

models

All available models are cartesian product of the following acoustic models and ctc decoders configurations.

Acoustic Models

1. Linear Layer
2. LSTM
3. DeepSpeech Toy
4. DeepSpeech Small
5. DeepSpeech Large

CTC Decoders

1. Greedy CTC
2. Lexicon CTC
3. LM CTC

you can evaluate all models using evaluate.py, or try a specific configuration with try.py

Training

Train command example:

python train.py --conf confs/linear_acoustic.json --logger wandb

Evaluation

The following command enables you to evaluate and observer the wer of all available models on train, test and val data

Eval command example:

python evaluate.py --conf confs/linear_acoustic.json

Try

A CLI which enables you to construct the ASR model you want and try to transcribe with it :)

Eval command example:

python try.py --conf confs/archive/try.json

Benchmarks

Evaluation results on all data are detailed in the paper. we appended here the graphs on the validation data.