This model is built upon Conformer architecture and trained using the CTC (Connectionist Temporal Classification) approach. The training dataset consists of 823 hours of French speech data sourced from the publicly available Common Voice 11.0.
- # of parameters (million): 89.98
- GPU info
- NVIDIA GeForce RTX 3090
- # of GPUs: 8
The script run.sh contains the overall model training process.
- Follow the steps
data_prep.mdand rundata_prep.shto prepare the datset and pronunciation lexicon for a given language. The second and fourth stages ofdata_prep.shinvolve language-specific special processing, which are detailed in thelang_process.md. - The training of this model utilized 8 NVIDIA GeForce RTX 3090 GPUs and took 21 hours. The detailed model parameters are detailed in
config.jsonandhyper-p.json. Dataset paths should be added to themetainfo.jsonfor efficient management of datasets.
-
To train the model:
`bash run.sh fr exp/Monolingual/fr --sta 1 --sto 3` -
To plot the training curves:
`python utils/plot_tb.py exp/Monolingual/fr/log/tensorboard/file -o exp/Monolingual/fr/monitor.png`
| Monitor figure |
|---|
 |
-
To decode with CTC and calculate the %PER:
`bash run.sh fr exp/Monolingual/fr --sta 4 --sto 4`test_fr %SER 56.81 | %PER 4.93 [ 28785 / 583591, 4987 ins, 7890 del, 15908 sub ]
-
For FST decoding,
config.jsonandhyper-p.jsonare needed to train language model. Notice the distinction between the profiles for training the ASR model and the profiles for training the language model, which have the same name but are in different directories. -
To decode with FST and calculate the %WER:
`bash run.sh fr exp/Monolingual/fr --sta 5`test_fr_ac0.9_lm0.8_wip0.0.hyp %SER 57.37 | %WER 15.58 [ 24216 / 155399, 2273 ins, 3021 del, 18922 sub ]
-
The files used to train this model and the trained model are available in the following table.
Pronunciation lexicon Checkpoint model Language model Tensorboard log lexicon_fr.txtMono_fr_best-3.ptlm_fr_4gram.arpatb_fr