This repository so far holds the recipes for ASR systems using the tedx data scraped from the web, as well as recipes segmenting the speech according to existing sentences, such as those produced by an MT aligner. The recipes are meant to be run inside of Kaldi.
- Make sure Kaldi is installed somewhere (KALDI_ROOT)
- Then in some directory run the following lines
git clone https://github.com/m-wiesner/tedx.git
mv tedx $KALDI_ROOT/egs/
- Go to the tedx s5 directory
cd KALDI_ROOT/egs/tedx/s5
- Train a basic acoustic model, decode a held-out evaluation set to compute WER, and align the MT data for language pairs specified in the text directory in run.sh:
./run.sh --src fr