|
--- |
|
language: |
|
- ja |
|
license: apache-2.0 |
|
tags: |
|
- generated_from_trainer |
|
datasets: |
|
- mozilla-foundation/common_voice_11_0 |
|
metrics: |
|
- wer |
|
- cer |
|
model-index: |
|
- name: wav2vec2-base-japanese-asr |
|
results: |
|
- task: |
|
type: automatic-speech-recognition |
|
name: Speech Recognition |
|
dataset: |
|
name: common_voice_11_0 |
|
type: common_voice |
|
args: ja |
|
metrics: |
|
- type: wer |
|
value: 14.177284 |
|
name: Test WER |
|
- type: cer |
|
value: 6.462501 |
|
name: Test CER |
|
- task: |
|
name: Speech Recognition |
|
type: automatic-speech-recognition |
|
dataset: |
|
name: Reazonspeech |
|
type: custom |
|
args: ja |
|
metrics: |
|
- name: Test WER |
|
type: wer |
|
value: 40.864413 |
|
- name: Test CER |
|
type: cer |
|
value: 29.367348 |
|
--- |
|
|
|
# wav2vec2-base-asr |
|
|
|
This model is a fine-tuned version of [rinna/japanese-wav2vec2-base](https://huggingface.co/rinna/japanese-wav2vec2-base) on the [common_voice_11_0 dataset](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0/viewer/ja) for ASR tasks. |
|
|
|
This model can only predict Hiragana. |
|
|
|
## Acknowledgments |
|
|
|
This model's fine-tuning approach was inspired by and references the training methodology used in [vumichien/wav2vec2-large-xlsr-japanese-hiragana](https://huggingface.co/vumichien/wav2vec2-large-xlsr-japanese-hiragana). |
|
|
|
## Training Procedure |
|
|
|
Fine-tuning on the common_voice_11_0 dataset led to the following results: |
|
|
|
| Step | Training Loss | Validation Loss | WER | |
|
|-------|---------------|-----------------|----------| |
|
| 1000 | 6.088100 | 3.452597 | 1.000000 | |
|
| 2000 | 2.816600 | 0.756278 | 0.263624 | |
|
| 3000 | 0.837600 | 0.471486 | 0.185915 | |
|
| 4000 | 0.624900 | 0.420854 | 0.159801 | |
|
| 5000 | 0.533300 | 0.392494 | 0.149141 | |
|
| 6000 | 0.490000 | 0.394669 | 0.144826 | |
|
| 7000 | 0.441600 | 0.379999 | 0.141807 | |
|
|
|
### Training hyperparameters |
|
|
|
The training hyperparameters remained consistent throughout the fine-tuning process: |
|
|
|
- learning_rate: 1e-4 |
|
- train_batch_size: 16 |
|
- eval_batch_size: 16 |
|
- seed: 42 |
|
- gradient_accumulation_steps: 2 |
|
- num_train_epochs: 20 |
|
- warmup_steps: 2000 |
|
- lr_scheduler_type: linear |
|
|
|
### How to evaluate the model |
|
|
|
```python |
|
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor |
|
from datasets import load_dataset |
|
import torch |
|
import torchaudio |
|
import librosa |
|
import numpy as np |
|
import re |
|
import MeCab |
|
import pykakasi |
|
from evaluate import load |
|
|
|
model = Wav2Vec2ForCTC.from_pretrained('TKU410410103/wav2vec2-base-japanese-asr') |
|
processor = Wav2Vec2Processor.from_pretrained("TKU410410103/wav2vec2-base-japanese-asr") |
|
|
|
# load dataset |
|
test_dataset = load_dataset('mozilla-foundation/common_voice_11_0', 'ja', split='test') |
|
remove_columns = [col for col in test_dataset.column_names if col not in ['audio', 'sentence']] |
|
test_dataset = test_dataset.remove_columns(remove_columns) |
|
|
|
# resample |
|
def process_waveforms(batch): |
|
speech_arrays = [] |
|
sampling_rates = [] |
|
|
|
for audio_path in batch['audio']: |
|
speech_array, _ = torchaudio.load(audio_path['path']) |
|
speech_array_resampled = librosa.resample(np.asarray(speech_array[0].numpy()), orig_sr=48000, target_sr=16000) |
|
speech_arrays.append(speech_array_resampled) |
|
sampling_rates.append(16000) |
|
|
|
batch["array"] = speech_arrays |
|
batch["sampling_rate"] = sampling_rates |
|
|
|
return batch |
|
|
|
# hiragana |
|
CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞", |
|
"؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]", |
|
"{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。", |
|
"、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽", |
|
"『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "'", "ʻ", "ˆ"] |
|
chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]" |
|
|
|
wakati = MeCab.Tagger("-Owakati") |
|
kakasi = pykakasi.kakasi() |
|
kakasi.setMode("J","H") |
|
kakasi.setMode("K","H") |
|
kakasi.setMode("r","Hepburn") |
|
conv = kakasi.getConverter() |
|
|
|
def prepare_char(batch): |
|
batch["sentence"] = conv.do(wakati.parse(batch["sentence"]).strip()) |
|
batch["sentence"] = re.sub(chars_to_ignore_regex,'', batch["sentence"]).strip() |
|
return batch |
|
|
|
|
|
resampled_eval_dataset = test_dataset.map(process_waveforms, batched=True, batch_size=50, num_proc=4) |
|
eval_dataset = resampled_eval_dataset.map(prepare_char, num_proc=4) |
|
|
|
# begin the evaluation process |
|
wer = load("wer") |
|
cer = load("cer") |
|
|
|
def evaluate(batch): |
|
inputs = processor(batch["array"], sampling_rate=16_000, return_tensors="pt", padding=True) |
|
with torch.no_grad(): |
|
logits = model(inputs.input_values.to(device), attention_mask=inputs.attention_mask.to(device)).logits |
|
pred_ids = torch.argmax(logits, dim=-1) |
|
batch["pred_strings"] = processor.batch_decode(pred_ids) |
|
return batch |
|
|
|
columns_to_remove = [column for column in eval_dataset.column_names if column != "sentence"] |
|
batch_size = 16 |
|
result = eval_dataset.map(evaluate, remove_columns=columns_to_remove, batched=True, batch_size=batch_size) |
|
|
|
wer_result = wer.compute(predictions=result["pred_strings"], references=result["sentence"]) |
|
cer_result = cer.compute(predictions=result["pred_strings"], references=result["sentence"]) |
|
|
|
print("WER: {:2f}%".format(100 * wer_result)) |
|
print("CER: {:2f}%".format(100 * cer_result)) |
|
``` |
|
|
|
### Test results |
|
The final model was evaluated as follows: |
|
|
|
On common_voice_11_0: |
|
- WER: 14.177284% |
|
- CER: 6.462501% |
|
|
|
On reazonspeech(tiny): |
|
- WER: 40.864413% |
|
- CER: 29.367348% |
|
### Framework versions |
|
|
|
- Transformers 4.39.1 |
|
- Pytorch 2.2.1+cu118 |
|
- Datasets 2.17.1 |
