It is consistent with my observation except that adding more layers up to 6 with 1024 units still obtains some improvement.

Thanks. Will follow your suggestions (when GPUs are available). Could you share your WER results? If it is better than https://github.com/espnet/espnet/blob/d51e76c0baa556e28a3e090335944478828fbc65/egs/librispeech/asr1/RESULTS, then, I may follow your network architecture, or ask you to make a PR.

The trained models can be provided through the release (e.g., https://github.com/espnet/espnet/releases/download/untagged-f5ccde023841a43380a9/librispeech_asr1.tgz)

No, my observation is from train_100. I think you system is the best til now.

@sw005320 Hi Shinji, for RWTH setup, is there any published paper for this? I couldn't find one online.

https://arxiv.org/pdf/1805.03294
This is not exactly same as what we're now using, but our set up is based on the discussion with them what would be most dominant given our current implementation.

From this paper, their system played with the subsampling factor 32 in the pertaining and 8 for fine-tuning and they showed some improvement.

In our babel-10 setting, we used subsampling factor 4, and we allowed the minimum number of frames in an utterance to be 10 frames.

Is there any specific reason use our current setting?

We don't have a specific reason, and we may test further subsampling like the RWTH paper, but I internally found that further subsampling (8) slightly degrades the performance in the Librispeech task.