Modèle Français 0.6
Pre-release
Pre-release
Jeux de données :
- Lingua Libre (~40h)
- Common Voice FR (v5.1) (~490h, en autorisant jusqu'à 32 duplicatas)
- Training Speech (~180h)
- African Accented French (~15h)
- M-AILABS French (~315h)
- Centre de Conférence Pierre Mendès France (~300h)
Total : ~1340h
Paramètres :
EPOCHS=32
LEARNING_RATE=0.0001
DROPOUT=0.3
BATCH_SIZE=64
LM_ALPHA=0.5919543900530122
LM_BETA=1.6082513974258137
Best params: lm_alpha=0.5919543900530122 and lm_beta=1.6082513974258137 with WER=0.29113864240896115
Language Model : dump wikipedia + dump débats assemblée nationale.
Licence : MPL 2.0 https://github.com/common-voice/commonvoice-fr/blob/5699e59244d14bb14d5b7603b91c934b761c9194/DeepSpeech/LICENSE.txt
Fonctionne avec DeepSpeech v0.7
, v0.8
, v0.9
.
Résultats test set:
Test on /mnt/extracted/data/African_Accented_French/African_Accented_French/African_Accented_French_test.csv - WER: 0.448976, CER: 0.242144, loss: 43.320114
Test on /mnt/extracted/data/M-AILABS/fr_FR/fr_FR_test.csv - WER: 0.097462, CER: 0.027961, loss: 12.057733
Test on /mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR_test.csv - WER: 0.199902, CER: 0.059519, loss: 15.992792
Test on /mnt/extracted/data/cv-fr/clips/test.csv - WER: 0.301279, CER: 0.142777, loss: 37.710129
Test on /mnt/extracted/data/lingualibre/lingua_libre_Q21-fra-French_test.csv - WER: 0.589222, CER: 0.182179, loss: 7.118075
Test on /mnt/extracted/data/ccpmf/transcriptionsXML_audioMP3_MEFR_CCPMF_2012-2020/ccpmf_test.csv - WER: 0.486848, CER: 0.304395, loss: 89.443710