rristo/icefall_conformer_ctc3_et

conformer-ctc et

Icefall conformer-ctc3 based recipe (https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conformer_ctc3) trained Estonian ASR model using ERR2020 dataset

• WER on ERR2020: 12.1
• WER on mozilla commonvoice_11: 23.2

For usage:

• clone this repo (git clone https://huggingface.co/rristo/icefall_conformer_ctc3_et)
• go to repo (cd icefall_conformer_ctc3_et)
• build docker image for needed libraries (build.sh or build.bat)
• run docker container (run.shor run.sh). This mounts current directory
• run notebook err2020/conformer_ctc3_usage.ipynb for example usage
  • currently expects audio to be in .wav format

Model description

ASR model for Estonian, uses Estonian Public Broadcasting data ERR2020 data (around 340 hours of audio)

Intended uses & limitations

Pretty much a toy model, trained on limited amount of data. Might not work well on data out of domain (especially spontaneous/noisy data).

Training and evaluation data

Trained on ERR2020 data, evaluated on ERR2020 and mozilla commonvoice test data.

Training procedure

Used Icefall conformer-ctc3 based recipe (https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conformer_ctc3)

Training results

TODO

Framework versions

• icefall
• k2
• kaldifeat==1.24
• lhotse==1.15.0
• torch==2.0.0