This model is for transcribing audio into Hiragana, one format of Japanese language.
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the mozilla-foundation/common_voice_8_0 dataset. Note that the following results are achieved by:
- Modify
eval.pyto suit the use case. - Since kanji and katakana shares the same sound as hiragana, we convert all texts to hiragana using pykakasi and tokenize them using fugashi.
It achieves the following results on the evaluation set:
- Loss: 0.7751
- Cer: 0.2227
Evaluation results (Running ./eval.py):
| Model | Metric | Common-Voice-8/test | speech-recognition-community-v2/dev-data |
|---|---|---|---|
| w/o LM | WER | 0.5964 | 0.5532 |
| CER | 0.2944 | 0.2629 | |
| w/ LM | WER | 0.5405 | 0.4877 |
| CER | 0.2754 | 0.2487 |
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: 5e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 1000
- training_steps: 4000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Cer |
|---|---|---|---|---|
| 4.4081 | 1.6 | 500 | 4.0983 | 1.0 |
| 3.303 | 3.19 | 1000 | 3.3563 | 1.0 |
| 3.1538 | 4.79 | 1500 | 3.2066 | 0.9239 |
| 2.1526 | 6.39 | 2000 | 1.1597 | 0.3355 |
| 1.8726 | 7.98 | 2500 | 0.9023 | 0.2505 |
| 1.7817 | 9.58 | 3000 | 0.8219 | 0.2334 |
| 1.7488 | 11.18 | 3500 | 0.7915 | 0.2222 |
| 1.7039 | 12.78 | 4000 | 0.7751 | 0.2227 |
| Stop & Train | ||||
| 1.6571 | 15.97 | 5000 | 0.6788 | 0.1685 |
| 1.520400 | 19.16 | 6000 | 0.6095 | 0.1409 |
| 1.448200 | 22.35 | 7000 | 0.5843 | 0.1430 |
| 1.385400 | 25.54 | 8000 | 0.5699 | 0.1263 |
| 1.354200 | 28.73 | 9000 | 0.5686 | 0.1219 |
| 1.331500 | 31.92 | 10000 | 0.5502 | 0.1144 |
| 1.290800 | 35.11 | 11000 | 0.5371 | 0.1140 |
| Stop & Train | ||||
| 1.235200 | 38.30 | 12000 | 0.5394 | 0.1106 |
Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2+cu102
- Datasets 1.18.2.dev0
- Tokenizers 0.11.0
