This repository contains scripts to train Jasper for German speech recognition.
Code was adapted from https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/SpeechRecognition/Jasper.
Code was copied from commit a2281e3.
To get training data follow instructions in https://github.com/ynop/megs. Download the LM from https://github.com/ynop/german-asr-lm.
Build docker container
scripts/docker/build.sh
Launch docker container
scripts/docker/launch.sh \
[german-asr-data]/data \
[output-dir]/checkpoints \
[output-dir]/results
Run training
scripts/train.sh
Word error rates in %, for megs v2.
| Decoding | Training-Data | dev | test |
|---|---|---|---|
| Best Path | train | 17.91 | 19.16 |
| Beam Search (6-gram LM) | train | 12.97 | 14.14 |
| Decoding | Training-Data | dev_cv | test_cv | dev_tuda | test_tuda |
|---|---|---|---|---|---|
| Best Path | train | 23.95 | 28.65 | 18.13 | 18.86 |
| Beam Search (6-gram LM) | train | 17.12 | 21.00 | 13.00 | 13.19 |
| Decoding | Training-Data | dev_swc | test_swc | dev_voxforge | test_voxforge |
|---|---|---|---|---|---|
| Best Path | train | 13.97 | 11.79 | 11.14 | 10.55 |
| Beam Search (6-gram LM) | train | 10.05 | 8.80 | 8.78 | 8.50 |