Releases: common-voice/commonvoice-fr
Releases · common-voice/commonvoice-fr
Modèle Français 0.6
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=32LEARNING_RATE=0.0001DROPOUT=0.3BATCH_SIZE=64LM_ALPHA=0.5919543900530122LM_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
Modèle Français 0.5.2
Jeux de données :
- Lingua Libre (~40h)
- Common Voice FR (v2) (~490h, en autorisant jusqu'à 32 duplicatas)
- Training Speech (~180h)
- African Accented French (~15h)
- M-AILABS French (~315h)
Total : ~1040h
Paramètres :
EPOCHS=30LEARNING_RATE=0.0001DROPOUT=0.3BATCH_SIZE=64LM_ALPHA=0.7203202402564637LM_BETA=1.5747698919871918
Language Model : dump wikipedia + dump débats assemblée nationale.
Fonctionne avec DeepSpeech v0.7, v0.8, v0.9.
Correction du packaging de kenlm.scorer
Correction des valeurs par défaut de alpha/beta dans kenlm.scorer
Résultats test set:
Test on /mnt/extracted/data/African_Accented_French/African_Accented_French/African_Accented_French_test.csv - WER: 0.442362, CER: 0.235577, loss: 42.941334
Test on /mnt/extracted/data/M-AILABS/fr_FR/fr_FR_test.csv - WER: 0.092794, CER: 0.026505, loss: 11.276774
Test on /mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR_test.csv - WER: 0.200373, CER: 0.059958, loss: 16.225618
Test on /mnt/extracted/data/cv-fr/clips/test.csv - WER: 0.300508, CER: 0.147202, loss: 39.204407
Test on /mnt/extracted/data/lingualibre/lingua_libre_Q21-fra-French_test.csv - WER: 0.577170, CER: 0.171211, loss: 6.977585
Modèle Français 0.5.1
Jeux de données :
- Lingua Libre (~40h)
- Common Voice FR (v2) (~490h, en autorisant jusqu'à 32 duplicatas)
- Training Speech (~180h)
- African Accented French (~15h)
- M-AILABS French (~315h)
Total : ~1040h
Paramètres :
EPOCHS=30LEARNING_RATE=0.0001DROPOUT=0.3BATCH_SIZE=64LM_ALPHA=0.7203202402564637LM_BETA=1.5747698919871918
Language Model : dump wikipedia + dump débats assemblée nationale.
Fonctionne avec DeepSpeech v0.7, v0.8.
Résultats test set:
Test on /mnt/extracted/data/African_Accented_French/African_Accented_French/African_Accented_French_test.csv - WER: 0.442362, CER: 0.235577, loss: 42.941334
Test on /mnt/extracted/data/M-AILABS/fr_FR/fr_FR_test.csv - WER: 0.092794, CER: 0.026505, loss: 11.276774
Test on /mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR_test.csv - WER: 0.200373, CER: 0.059958, loss: 16.225618
Test on /mnt/extracted/data/cv-fr/clips/test.csv - WER: 0.300508, CER: 0.147202, loss: 39.204407
Test on /mnt/extracted/data/lingualibre/lingua_libre_Q21-fra-French_test.csv - WER: 0.577170, CER: 0.171211, loss: 6.977585
Modèle Français 0.4
Jeux de données :
- Lingua Libre (~20h)
- Common Voice FR (v2) (~290h, en autorisant jusqu'à 8 duplicatas)
- Training Speech (~180h)
- African Accented French (~15h)
- M-AILABS French (~315h)
Total : ~820h
Paramètres :
- LEARNING_RATE=0.0001
- DROPOUT=0.3
- BATCH_SIZE=64
- LM_ALPHA=0.65
- LM_BETA=1.45
Language Model : dump wikipedia + dump débats assemblée nationale.
Fonctionne avec DeepSpeech v0.6.1.
Résultats test set:
Test on /mnt/extracted/data/lingualibre/lingua_libre_Q21-fra-French_test.csv - WER: 0.541340, CER: 0.150946, loss: 5.962852
--------------------------------------------------------------------------------
WER: 5.000000, CER: 0.241379, loss: 3.496368
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Lyokoï/électroencéphalographiquement.wav
- src: "électroencéphalographiquement"
- res: "électro en céphale orphique ment"
--------------------------------------------------------------------------------
WER: 4.000000, CER: 0.333333, loss: 3.654961
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Lyokoï/aposématisme.wav
- src: "aposématisme"
- res: "a posé ma time"
--------------------------------------------------------------------------------
WER: 4.000000, CER: 0.400000, loss: 4.680493
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/oligoasthénotératospermie.wav
- src: "oligoasthénotératospermie"
- res: "aligoté notera to sperm"
--------------------------------------------------------------------------------
WER: 4.000000, CER: 0.285714, loss: 7.043005
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/octingentesimo.wav
- src: "octingentesimo"
- res: "acting en tesi mo"
--------------------------------------------------------------------------------
WER: 4.000000, CER: 0.500000, loss: 12.178319
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Lyokoï/limousinerie.wav
- src: "limousinerie"
- res: "il vous i neri"
--------------------------------------------------------------------------------
WER: 4.000000, CER: 0.263158, loss: 17.644501
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/paléontologiquement.wav
- src: "paléontologiquement"
- res: "pale on a logiquement"
--------------------------------------------------------------------------------
WER: 4.000000, CER: 0.538462, loss: 20.121408
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/mielleusement.wav
- src: "mielleusement"
- res: "in a le cement"
--------------------------------------------------------------------------------
WER: 4.000000, CER: 0.454545, loss: 23.273678
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Poslovitch/ennuagement.wav
- src: "ennuagement"
- res: "en eut age ment"
--------------------------------------------------------------------------------
WER: 4.000000, CER: 0.692308, loss: 36.408180
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Xenophôn/Hondevilliers.wav
- src: "hondevilliers"
- res: "on ne vit le"
--------------------------------------------------------------------------------
WER: 4.000000, CER: 0.687500, loss: 38.046669
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/téléconsultation.wav
- src: "téléconsultation"
- res: "tel que les consultations"
--------------------------------------------------------------------------------
Test on /mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR_test.csv - WER: 0.197745, CER: 0.059797, loss: 17.292450
--------------------------------------------------------------------------------
WER: 4.000000, CER: 1.333333, loss: 38.737186
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/LeComteDeMonteCristoT1Chap5_0237.converted.wav
- src: "espoir"
- res: "n est ce soir"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 1.000000, loss: 47.523190
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/MonsieurLecoqP1C16_0188.converted.wav
- src: "continuez"
- res: "quand il est"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.250000, loss: 0.010373
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/MonsieurLecoqT2P16_0185.converted.wav
- src: "chanlouineau"
- res: "chan luneau"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.142857, loss: 0.052286
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/MonsieurLecoqP1C42_0070.converted.wav
- src: "parbleu"
- res: "par bleu"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.142857, loss: 0.219133
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/LeComteDeMonteCristoT1Chap3_0284.converted.wav
- src: "pardieu"
- res: "par dieu"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.333333, loss: 1.239774
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/LesMysteresDeParisT3P5C14_0002.converted.wav
- src: "amitie"
- res: "a miti"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.384615, loss: 1.923999
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/LeComteDeMonteCristoT1Chap24_0002.converted.wav
- src: "eblouissement"
- res: "et boisement"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.250000, loss: 2.610425
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/MonsieurLecoqT2P33_0032.converted.wav
- src: "chimeres"
- res: "chi mere"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.500000, loss: 3.350882
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/MonsieurLecoqT2P04_0012.converted.wav
- src: "hola"
- res: "a la"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.400000, loss: 7.205533
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/LeDernierJourDunCondamne_0712.converted.wav
- src: "lirlonfa malure"
- res: "le lan fan maure"
--------------------------------------------------------------------------------
Test on /mnt/extracted/data/M-AILABS/fr_FR/fr_FR_test.csv - WER: 0.090398, CER: 0.025351, loss: 11.177062
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.166667, loss: 3.342017
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/female/ezwa/monsieur_lecoq/wavs/monsieur_lecoq_2_36_f000179.wav
- src: "dubois"
- res: "du bois"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.857143, loss: 8.253085
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/female/nadine_eckert_boulet/les_tribulations_dun_chinoise/wavs/les_tribulations_dun_chinoise_10_f000043.wav
- src: "bidulph"
- res: "le bip"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.375000, loss: 10.294103
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/male/gilles_g_le_blanc/lupin_contre_holmes/wavs/lupin_contre_holmes_13_f000184.wav
- src: "personne"
- res: "le songe"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 6.000000, loss: 20.541677
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/female/nadine_eckert_boulet/les_mysteres_de_paris/wavs/les_mysteres_de_paris_4_13_f000027.wav
- src: "m"
- res: "on ne "
--------------------------------------------------------------------------------
WER: 1.500000, CER: 0.400000, loss: 4.110573
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/male/gilles_g_le_blanc/lupin_contre_holmes/wavs/lupin_contre_holmes_07_f000165.wav
- src: "m destange"
- res: "mais des tange"
--------------------------------------------------------------------------------
WER: 1.500000, CER: 0.266667, loss: 4.140529
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/male/gilles_g_le_blanc/lupin_contre_holmes/wavs/lupin_contre_holmes_14_f000218.wav
- src: "langlais ricana"
- res: "l'anglais et cana"
--------------------------------------------------------------------------------
WER: 1.200000, CER: 0.279070, loss: 58.677330
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/female/ezwa/monsieur_lecoq/wavs/monsieur_lecoq_2_40_f000027.wav
- src: "incompréhensible balbutia t il inimaginable"
- res: "un coupé aussi ble balbutiant il imaginable"
--------------------------------------------------------------------------------
WER: 1.000000, CER: 0.125000, loss: 0.046964
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/male/gilles_g_le_blanc/lupin_contre_holmes/wavs/lupin_contre_holmes_11_f000012.wav
- src: "ganimard"
- res: "gaimard"
--------------------------------------------------------------------------------
WER: 1.000000, CER: 0.142857, loss: 0.094500
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/male/gilles_g_le_blanc/lupin_contre_holmes/wavs/lupin_contre_holmes_01_f000115.wav
- src: "gerbois"
- res: "gerboise"
--------------------------------------------------------------------------------
WER: 1.000000, CER: 0.150000, loss: 0.097039
- wav: file:///mnt/extracted/data/M-AILABS/fr_FR/female/ezwa/monsieur_lecoq/wavs/monsieur_lecoq_2_49_f000013.wav
- src: "chanlouineau fusillé"
- res: "chanoine au fusillé"
--------------------------------------------------------------------------------
Test on /mnt/extracted/data/African_A...
Modèle Français 0.3.4
Jeux de données :
- Lingua Libre (~20h)
- Common Voice FR (v2) (~120h, en autorisant des duplicatas)
- Training Speech (~180h)
- African Accented French (~15h)
- M-AILABS French (~315h)
Total : ~650h
Paramètres :
- LEARNING_RATE=0.0001
- DROPOUT=0.3
- BATCH_SIZE=96
- LM_ALPHA=0.65
- LM_BETA=1.45
Language Model : dump wikipedia + dump débats assemblée nationale.
Fonctionne avec DeepSpeech v0.6.0. Ré-export de 0.3.3 pour corriger un bug dans TFLite
Résultats test set:
Testing model on /mnt/extracted/data/lingualibre/lingua_libre_Q21-fra-French_test.csv
Test epoch | Steps: 75 | Elapsed Time: 0:01:44
Test on /mnt/extracted/data/lingualibre/lingua_libre_Q21-fra-French_test.csv - WER: 0.467659, CER: 0.138508, loss: 6.800947
--------------------------------------------------------------------------------
WER: 4.000000, CER: 2.200000, loss: 39.604939
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Vahidmasrour/abhal.wav
- src: "abhal"
- res: "le panel a bal"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 0.600000, loss: 2.462182
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/irato.wav
- src: "irato"
- res: "il a to"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 0.111111, loss: 3.428576
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/ultratrifoliophile.wav
- src: "ultratrifoliophile"
- res: "ultra trifolio phile"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 0.333333, loss: 5.036440
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/cuthomiurophile.wav
- src: "cuthomiurophile"
- res: "culto miro phile"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 0.333333, loss: 5.090287
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/remiauler.wav
- src: "remiauler"
- res: "remi au le"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 0.285714, loss: 6.972348
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Lyokoï/indoeuropéiste.wav
- src: "indoeuropéiste"
- res: "in doro péiste"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 0.454545, loss: 7.742430
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Jules78120/Antarctique.wav
- src: "antarctique"
- res: "en parti que"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 0.833333, loss: 8.499911
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Guilhelma (Ives)/padena.wav
- src: "padena"
- res: "pas de nom"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 0.307692, loss: 8.974085
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/Lyokoï/pleurogynique.wav
- src: "pleurogynique"
- res: "pleu rogi mique"
--------------------------------------------------------------------------------
WER: 3.000000, CER: 0.230769, loss: 9.156916
- wav: file:///mnt/extracted/data/lingualibre/lingua_libre/Q21-fra-French/WikiLucas00/bonhomisation.wav
- src: "bonhomisation"
- res: "bon ami sation"
--------------------------------------------------------------------------------
Testing model on /mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR_test.csv
Test epoch | Steps: 129 | Elapsed Time: 0:10:36
Test on /mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR_test.csv - WER: 0.185852, CER: 0.061034, loss: 21.406639
--------------------------------------------------------------------------------
WER: 4.000000, CER: 1.222222, loss: 28.107866
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/MonsieurLecoqP1C16_0188.converted.wav
- src: "continuez"
- res: "quand il ne est"
--------------------------------------------------------------------------------
WER: 2.333333, CER: 0.818182, loss: 123.491005
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/LesMysteresDeParisT1P1C5_0129.converted.wav
- src: "diminution de fourloir"
- res: "des minutions de de fournoue a sa fin"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.142857, loss: 0.981466
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/LaGloireDuComacchio_0097.converted.wav
- src: "pardieu"
- res: "par dieu"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.500000, loss: 5.709780
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/LeComteDeMonteCristoT1Chap3_0240.converted.wav
- src: "hola"
- res: "a la"
--------------------------------------------------------------------------------
WER: 2.000000, CER: 0.750000, loss: 6.360806
- wav: file:///mnt/extracted/data/trainingspeech/ts_2019-04-11_fr_FR/madamebovaryC24_0123.converted.wav
- src: "leon"
- res: "et on"
---------------------------...
Modèle Français 0.3.2
Jeux de données :
- Lingua Libre (~20h)
- Common Voice FR (v2) (~120h, en autorisant des duplicatas)
- Training Speech (~180h)
- African Accented French (~15h)
- M-AILABS French (~315h)
Total : ~650h
Paramètres :
- LEARNING_RATE=0.0001
- DROPOUT=0.3
- BATCH_SIZE=64
- LM_ALPHA=0.65
- LM_BETA=1.4
Language Model : dump wikipedia + dump débats assemblée nationale.
Fonctionne avec DeepSpeech v0.6.0-alpha.15.
Modèle Français 0.3.1
Jeux de données :
- Lingua Libre (~20h)
- Common Voice FR (v2) (~120h, en autorisant des duplicatas)
- Training Speech (~180h)
- African Accented French (~15h)
- M-AILABS French (~315h)
Total : ~650h
Paramètres :
- LEARNING_RATE=0.0001
- DROPOUT=0.3
- BATCH_SIZE=64
- LM_ALPHA=0.65
- LM_BETA=1.4
Language Model : dump wikipedia + dump débats assemblée nationale.
Fonctionne avec DeepSpeech v0.6.0-alpha.14.
Modèle Français 0.3
Jeux de données :
- Lingua Libre (~20h)
- Common Voice FR (v2) (~120h, en autorisant des duplicatas)
- Training Speech (~180h)
- African Accented French (~15h)
- M-AILABS French (~315h)
Total : ~650h
Paramètres :
- LEARNING_RATE=0.0001
- DROPOUT=0.3
- BATCH_SIZE=64
- LM_ALPHA=0.65
- LM_BETA=1.4
Language Model : dump wikipedia + dump débats assemblée nationale.
Fonctionne avec DeepSpeech v0.6.0-alpha.10.
Modèle Français 0.2
Seconde version alpha d'un modèle français.
- LEARNING_RATE=0.0001
- DROPOUT=0.2
- LM_ALPHA=0.85
- LM_BETA=1.45
Discussions et retours sur https://discourse.mozilla.org/t/modele-francais-0-2-pour-deepspeech-v0-6/45808
Modèle Français 0.1
Première version alpha d'un modèle français.
Merci de bien lire le message descriptif sur Discourse: https://discourse.mozilla.org/t/un-premier-modele-francais/41100
Les retours sont les bienvenus.