Speech to text with catalyst

Data used for testing:

• asr_public_stories_2 mp3+txt from this repo.

How to run

Create shell file with structure like this:

#!/bin/bash

CONF=<path to config file>
LOGDIR=<path to log dir>
# num timesteps before model == num timesteps after model
export MODEL_OUTPUTS_LENGTH=0
# if do not need to inform about training process please remove
# appropriate callback from configuration file
export CATALYST_TELEGRAM_TOKEN="<bot token>"
export CATALYST_TELEGRAM_CHAT_ID="<chat id>"

[ -e ${LOGDIR} ] && rm -rf ${LOGDIR} && echo " * Removed already existed logdir - '${LOGDIR}'"
catalyst-dl run --expdir sound --config ${CONF} --logdir ${LOGDIR} --verbose

Implemented models:

• Deep Speech: Scaling up end-to-end speech recognition
• Deep Speech 2: End-to-End Speech Recognition in English and Mandarin
• Lookahead Convolution Layer For Unidirectional Recurrent Neural Networks