đ XLS-R-1B - French
This is a fine - tuned model for automatic speech recognition, leveraging the power of the pre - trained facebook/wav2vec2-xls-r-1b on the French dataset MOZILLA - FOUNDATION/COMMON_VOICE_8_0. It offers high - quality speech recognition capabilities with excellent performance metrics.
⨠Features
- Automatic Speech Recognition: Specialized for French speech recognition tasks.
- Fine - Tuned on Quality Data: Trained on the MOZILLA - FOUNDATION/COMMON_VOICE_8_0 French dataset.
- Multiple Evaluation Metrics: Evaluated using WER and CER, both with and without a language model.
đĻ Installation
No installation steps provided in the original document, so this section is skipped.
đģ Usage Examples
No usage code examples provided in the original document, so this section is skipped.
đ Documentation
Model description
This model is a fine - tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA - FOUNDATION/COMMON_VOICE_8_0 - FR dataset.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 7.5e - 05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- gradient_accumulation_steps: 8
- total_train_batch_size: 128
- optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 2000
- num_epochs: 6.0
- mixed_precision_training: Native AMP
Training results
Training Loss |
Epoch |
Step |
Validation Loss |
Wer |
0.9827 |
0.29 |
1000 |
inf |
0.2937 |
1.0203 |
0.57 |
2000 |
inf |
0.2711 |
1.0048 |
0.86 |
3000 |
inf |
0.2620 |
0.9858 |
1.15 |
4000 |
inf |
0.2522 |
0.9709 |
1.43 |
5000 |
inf |
0.2365 |
0.9347 |
1.72 |
6000 |
inf |
0.2332 |
0.9256 |
2.01 |
7000 |
inf |
0.2261 |
0.8936 |
2.29 |
8000 |
inf |
0.2203 |
0.877 |
2.58 |
9000 |
inf |
0.2096 |
0.8393 |
2.87 |
10000 |
inf |
0.2017 |
0.8156 |
3.15 |
11000 |
inf |
0.1936 |
0.8015 |
3.44 |
12000 |
inf |
0.1880 |
0.774 |
3.73 |
13000 |
inf |
0.1834 |
0.8372 |
4.01 |
14000 |
inf |
0.1934 |
0.8075 |
4.3 |
15000 |
inf |
0.1923 |
0.8069 |
4.59 |
16000 |
inf |
0.1877 |
0.8064 |
4.87 |
17000 |
inf |
0.1955 |
0.801 |
5.16 |
18000 |
inf |
0.1891 |
0.8022 |
5.45 |
19000 |
inf |
0.1895 |
0.792 |
5.73 |
20000 |
inf |
0.1854 |
It achieves the best result on the validation set on STEP 13000:
Some problem occurs when calculating the validation loss.
Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2+cu102
- Datasets 1.18.3.dev0
- Tokenizers 0.11.0
Evaluation Commands
- To evaluate on
mozilla - foundation/common_voice_8
with split test
python eval.py --model_id Plim/xls - r - 1b - cv_8 - fr --dataset mozilla - foundation/common_voice_8_0 --config fr --split test
- To evaluate on
speech - recognition - community - v2/dev_data
python eval.py --model_id Plim/xls - r - 1b - cv_8 - fr --dataset speech - recognition - community - v2/dev_data --config fr --split validation --chunk_length_s 5.0 --stride_length_s 1.0
Evaluation Results
Without LM:
Dataset |
WER |
CER |
TEST CV |
18.33 |
5.60 |
DEV audio |
31.33 |
13.20 |
TEST audio |
/ |
/ |
With LM:
Dataset |
WER |
CER |
TEST CV |
15.40 |
5.36 |
DEV audio |
25.05 |
12.45 |
TEST audio |
/ |
/ |
đ§ Technical Details
Model Index
Property |
Details |
Model Type |
Fine - tuned version of facebook/wav2vec2-xls-r-1b for French automatic speech recognition |
Training Data |
MOZILLA - FOUNDATION/COMMON_VOICE_8_0 - FR dataset |
Task Results
The model has been evaluated on multiple tasks with different datasets:
- Automatic Speech Recognition on Common Voice 8:
- Test WER (with LM): 15.4
- Test CER (with LM): 5.36
- Automatic Speech Recognition on Robust Speech Event - Dev Data:
- Test WER (with LM): 25.05
- Test CER (with LM): 12.45
- Automatic Speech Recognition on Robust Speech Event - Test Data:
đ License
This project is licensed under the Apache - 2.0 license.