This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - JA dataset.
Kanji are converted into Hiragana using the pykakasi library during training and evaluation. The model can output both Hiragana and Katakana characters. Since there is no spacing, WER is not a suitable metric for evaluating performance and CER is more suitable.
On mozilla-foundation/common_voice_8_0 it achieved:
- cer: 23.64%
On speech-recognition-community-v2/dev_data it achieved:
- cer: 30.99%
It achieves the following results on the evaluation set:
- Loss: 0.5212
- Wer: 1.3068
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 7.5e-05
- train_batch_size: 48
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 2000
- num_epochs: 50.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 4.0974 | 4.72 | 1000 | 4.0178 | 1.9535 |
| 2.1276 | 9.43 | 2000 | 0.9301 | 1.2128 |
| 1.7622 | 14.15 | 3000 | 0.7103 | 1.5527 |
| 1.6397 | 18.87 | 4000 | 0.6729 | 1.4269 |
| 1.5468 | 23.58 | 5000 | 0.6087 | 1.2497 |
| 1.4885 | 28.3 | 6000 | 0.5786 | 1.3222 |
| 1.451 | 33.02 | 7000 | 0.5726 | 1.3768 |
| 1.3912 | 37.74 | 8000 | 0.5518 | 1.2497 |
| 1.3617 | 42.45 | 9000 | 0.5352 | 1.2694 |
| 1.3113 | 47.17 | 10000 | 0.5228 | 1.2781 |
Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2+cu102
- Datasets 1.18.2.dev0
- Tokenizers 0.11.0
Evaluation Commands
- To evaluate on
mozilla-foundation/common_voice_8_0with splittest
python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-300m-japanese --dataset mozilla-foundation/common_voice_8_0 --config ja --split test --log_outputs
- To evaluate on
mozilla-foundation/common_voice_8_0with splittest
python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-300m-japanese --dataset speech-recognition-community-v2/dev_data --config de --split validation --chunk_length_s 5.0 --stride_length_s 1.0