German Speech Recognition using Kaldi

Scripts to train Kaldi model for German speech recognition.

Data / LM / Lexicon

First, we have to get the data, a language model and the lexicon.

• To get the data follow the steps in https://github.com/ynop/megs.
• Download the LM from https://github.com/ynop/german-asr-lm.
• Download the lexicon from https://github.com/ynop/german-asr-lexicon.

Preparation

Before training, preparation of data, lexicon and lm has to be done by executing the script prepare.sh. In order to do that some python dependencies have to be installed with pip install -r requirements.

./prepare.sh \
    [german-asr-data]/data/full_waverized \
    [lexicon] \
    [sequitur-model] \
    [lm]

Training

After preparation, the actual training is done. At this step kaldi is used. To run it the easiest was is to used the docker image from https://hub.docker.com/r/kaldiasr/kaldi. All commands are in run.sh. This script is derived from the LibriSpeech recipe at egs/librispeech.

Results

Word error rates in %, for megs v2.

Model	Training-Data	dev	test
tdnn-chain	train	14.12	15.42

Model	Training-Data	dev_cv	test_cv	dev_tuda	test_tuda
tdnn-chain	train	14.71	18.45	11.85	12.80

Model	Training-Data	dev_swc	test_swc	dev_voxforge	test_voxforge
tdnn-chain	train	18.74	17.45	7.78	8.25