Training a LSTM for ASR

[stefan-falk]

I intend to try out an LSTM for speech recognition. Looking at the t2t code I noticed that there's a lstm_asr_v1 hparams-set which is should probably work with a lstm_seq2seq_attention?

However, it's not clear to me how the data has to be presented. At the moment I am using just a regular ASR dataset, similar to what we'd use when training a transformer model, but this does not seem to work right away with the LSTM model.

Is there any example for a ASR-LSTM that I can take a look at?

https://github.com/tensorflow/tensor2tensor/blob/a9da9635917814af890a31a060c5b29d31b2f906/tensor2tensor/models/lstm.py#L361

Update

I notice that using model lstm_seq2seq with lstm_asr_v1 does seem to "work". The training is starting. I don't know yet whether something useful will emerge but it's not crashing at least.

Only the lstm_seq2seq_attention does seem to have a problem with the provided input.

[snpushpi]

Hello , I was trying to write the same thing, may I know what did you use for checkpoint name, I mean, ckpt_name ? TIA!

[stefan-falk]

@snpushpi I didn't touch that. I guess it's the usual checkpoint name ckpt-xxxxx.