search

mravanelli / pytorch-kaldi

pytorch-kaldi is a project for developing state-of-the-art DNN/RNN hybrid speech recognition systems. The DNN part is managed by pytorch, while feature extraction, label computation, and decoding are performed with the kaldi toolkit.
2.37k stars 446 forks source link

Use final_architecture1.pkl for live test #255

Closed spacecd closed 3 years ago

spacecd commented 3 years ago

Hi Sir, Thanks for the detailed README.md, I managed to finish model training following Librispeech tutorial using libri_MLP_fmllr.cfg and got the result: %WER 9.54 [ 5018 / 52576, 558 ins, 643 del, 3817 sub ] /pytorch-kaldi/repo/exp_torch/libri_MLP_fmllrGPU c75_batch256/decode_test_clean_out_dnn1/wer_13_0.0,

I’d like to move the trained model exp_files/final_architecture1.pkl to one of our hardware recognition system for live test. Since our existing system loads final.mdl in Kaldi format, does anyone can give any idea for converting final_architecture1.pkl to Kaldi format? Thanks in advance :-D

mravanelli commented 3 years ago

Hi, we aren't supporting pytorch-kaldi anymore because we now haveca much better toolkit called speechbrain. My suggestion is to switch to it.

On Sun, Aug 1, 2021, 11:59 PM spacecd @.***> wrote:

Hi Sir, Thanks for the detailed README.md, I managed to finish model training following Librispeech tutorial using libri_MLP_fmllr.cfg and got the result: %WER 9.54 [ 5018 / 52576, 558 ins, 643 del, 3817 sub ] /pytorch-kaldi/repo/exp_torch/libri_MLP_fmllrGPU c75_batch256/decode_test_clean_out_dnn1/wer_13_0.0,

I’d like to move the trained model exp_files/final_architecture1.pkl to one of our hardware recognition system for live test. Since our existing system loads final.mdl in Kaldi format, does anyone can give any idea for converting final_architecture1.pkl to Kaldi format? Thanks in advance :-D

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/mravanelli/pytorch-kaldi/issues/255, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEA2ZVX7RSXB5XM4IJ3XGK3T2YJYLANCNFSM5BL4QHRA .

spacecd commented 3 years ago

Thanks Mirco for the reply. I’ve heard the name of SpeechBrain in the community, and will start follow it :-D

A quick question, is it possible that we convert the ASR model trained by SpeechBrain to Kaldi format? I ask this because our existing hardware recognition system is running C optimized kaldi decoder.

Hi, we aren't supporting pytorch-kaldi anymore because we now haveca much better toolkit called speechbrain. My suggestion is to switch to it. … On Sun, Aug 1, 2021, 11:59 PM spacecd @.***> wrote: Hi Sir, Thanks for the detailed README.md, I managed to finish model training following Librispeech tutorial using libri_MLP_fmllr.cfg and got the result: %WER 9.54 [ 5018 / 52576, 558 ins, 643 del, 3817 sub ] /pytorch-kaldi/repo/exp_torch/libri_MLP_fmllrGPU c75_batch256/decode_test_clean_out_dnn1/wer_13_0.0, I’d like to move the trained model exp_files/final_architecture1.pkl to one of our hardware recognition system for live test. Since our existing system loads final.mdl in Kaldi format, does anyone can give any idea for converting final_architecture1.pkl to Kaldi format? Thanks in advance :-D — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#255>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEA2ZVX7RSXB5XM4IJ3XGK3T2YJYLANCNFSM5BL4QHRA .

mravanelli commented 3 years ago

Unfortunately this is not possible..

On Mon, Aug 2, 2021, 1:22 AM spacecd @.***> wrote:

Thanks Mirco for the reply. I’ve heard the name of SpeechBrain in the community, and will start follow it :-D

A quick question, is it possible that we convert the ASR model trained by SpeechBrain to Kaldi format? I ask this because our existing hardware recognition system is running C optimized kaldi decoder.

Hi, we aren't supporting pytorch-kaldi anymore because we now haveca much better toolkit called speechbrain. My suggestion is to switch to it. … <#m-4930076304115183559> On Sun, Aug 1, 2021, 11:59 PM spacecd @.***> wrote: Hi Sir, Thanks for the detailed README.md, I managed to finish model training following Librispeech tutorial using libri_MLP_fmllr.cfg and got the result: %WER 9.54 [ 5018 / 52576, 558 ins, 643 del, 3817 sub ] /pytorch-kaldi/repo/exp_torch/libri_MLP_fmllrGPU c75_batch256/decode_test_clean_out_dnn1/wer_13_0.0, I’d like to move the trained model exp_files/final_architecture1.pkl to one of our hardware recognition system for live test. Since our existing system loads final.mdl in Kaldi format, does anyone can give any idea for converting final_architecture1.pkl to Kaldi format? Thanks in advance :-D — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#255 https://github.com/mravanelli/pytorch-kaldi/issues/255>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEA2ZVX7RSXB5XM4IJ3XGK3T2YJYLANCNFSM5BL4QHRA .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/mravanelli/pytorch-kaldi/issues/255#issuecomment-890725355, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEA2ZVWYWQDWLNCXWMCHYPLT2YTQLANCNFSM5BL4QHRA .

spacecd commented 3 years ago

Thanks Micro for your kindly help