search

gooofy / zamia-speech

Open tools and data for cloudless automatic speech recognition
GNU Lesser General Public License v3.0
443 stars 86 forks source link

Initial support for French #54

Closed pguyot closed 5 years ago

pguyot commented 5 years ago

Initial support for French with IPA lexicon and cleaned transcripts for Voxforge and M AILABS corpora. Also includes instructions to use Est Republicain corpus to train a language model.

svenha commented 5 years ago

Good progress! Do you have any WER results for French already?

pguyot commented 5 years ago

Not yet. Voxforge and M AILABS are the first two corpora for which I reviewed the transcripts, and through this review I adapted the tokenizer and created this large IPA lexicon. I mean to also include Mozilla Common Voice and, more interestingly for my purposes, TCOF which include children/adults conversations.

pguyot commented 5 years ago

Eventually finished a tdnn_250 model based on two corpora. On a Tesla T4, learning rate was about 1 iteration per 50 seconds.

%WER 17.40 [ 5018 / 28843, 356 ins, 755 del, 3907 sub ] exp/nnet3_chain/tdnn_250/decode_test/wer_9_0.0

gooofy commented 5 years ago

congratulations on your first french model! :) are you planning to run a tdnn_f model / adapt to larger LMs as well? would be interesting to see how much WER improves.

do you have model stats? how many hours of french training material, dict size etc?