tensorflow / lingvo

Lingvo
Apache License 2.0
2.82k stars 445 forks source link

ASR training recipe for inexperienced users #222

Open q121q opened 4 years ago

q121q commented 4 years ago

Hello,

I want to train an ASR system using the specaugment idea, however I have low experience with tensorflow & lingvo. What is the proposed way to move forward with this? From what I looked there's no end-to-end recipe, is there any other repo that contains such recipe?

Thanks!

drpngx commented 4 years ago

Alex, I would suggest:

  1. Make sure you can run the librispeech ASR recipe. This follows the same template as the MT recipe. You do that by:
    • Installing the package: pip install lingvo or from source.
    • Running the scripts in lingvo/tasks/asr/tools in order (librispeech.01.download_train.sh, librispeech.02.download_devtest.sh and so forth.
    • Running the trainer: bazel-bin/lingvo/trainer --run_locally=cpu --mode=sync --model=librispeech.Librispeech960Wpm --decoder_jobs=dev,test --job=controller,trainer_client,decoder_test,decoder_dev,evaler_test,evaler_dev --logdir=/tmp/logdir --logtostderr. See the MT example for a more elaborate scenario.
  2. Derive from the Librispeech config.
  3. Set params.encoder.use_specaugment = True in Task() (see the definition). You can further configure it using specaugment_network.