First set of pre-trained models for AN4 and for Librispeech. Two models are given; a smaller/lightweight LSTM version called DeepSpeech-light for Librispeech, as well as a pure DeepSpeech model (useful if you plan on adding more data, don't mind the additional memory usage). To download these, look at attached binaries to this release.
Both are CUDA only models since they use the cudnn library for the RNNs.
See the README for information on how to load these models and start re-training.
AN4
an4Test
| Network | WER | CER |
|---|---|---|
| DeepSpeech-light | N/A | N/A |
| DeepSpeech | 12 | 3.07 |
LibriSpeech
Librispeech-test-clean
| Network | WER | CER |
|---|---|---|
| DeepSpeech-light | 15 | 1.34 |
| DeepSpeech | 12 | 1.55 |
Librispeech-test-other
| Network | WER | CER |
|---|---|---|
| DeepSpeech-light | 36 | 3.80 |
| DeepSpeech | 33 | 3.24 |
I will re-train the lightweight model to get lower WERs on Librispeech for the next release hopefully!