anton-l's picture
anton-l HF staff
Upload README.md
831b85e
|
raw
history blame
4.39 kB
---
language:
- ja
license: apache-2.0
tags:
- automatic-speech-recognition
- generated_from_trainer
- hf-asr-leaderboard
- ja
- mozilla-foundation/common_voice_8_0
- robust-speech-event
datasets:
- mozilla-foundation/common_voice_8_0
model-index:
- name: XLS-R-300-m
results:
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: Common Voice 8
type: mozilla-foundation/common_voice_8_0
args: ja
metrics:
- name: Test WER
type: wer
value: 95.82
- name: Test CER
type: cer
value: 23.64
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: Robust Speech Event - Dev Data
type: speech-recognition-community-v2/dev_data
args: de
metrics:
- name: Test WER
type: wer
value: 100.0
- name: Test CER
type: cer
value: 30.99
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: Robust Speech Event - Dev Data
type: speech-recognition-community-v2/dev_data
args: ja
metrics:
- name: Test CER
type: cer
value: 30.37
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: Robust Speech Event - Test Data
type: speech-recognition-community-v2/eval_data
args: ja
metrics:
- name: Test CER
type: cer
value: 34.42
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
#
This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - JA dataset.
Kanji are converted into Hiragana using the [pykakasi](https://pykakasi.readthedocs.io/en/latest/index.html) 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
1. To evaluate on `mozilla-foundation/common_voice_8_0` with split `test`
```bash
python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-300m-japanese --dataset mozilla-foundation/common_voice_8_0 --config ja --split test --log_outputs
```
2. To evaluate on `mozilla-foundation/common_voice_8_0` with split `test`
```bash
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
```