TKU410410103's picture
Update README.md
9206637 verified
metadata
license: apache-2.0
tags:
  - generated_from_trainer
metrics:
  - wer
  - cer
model-index:
  - name: hubert-base-japanese-asr
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: common_voice_11_0
          type: common_voice
          args: ja
        metrics:
          - name: Test WER
            type: wer
            value: 27.511982
          - name: Test CER
            type: cer
            value: 11.699897
datasets:
  - mozilla-foundation/common_voice_11_0
language:
  - ja

hubert-base-asr

This model is a fine-tuned version of rinna/japanese-hubert-base on the common_voice_11_0 dataset 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.

Training Procedure

Fine-tuning on the common_voice_11_0 dataset led to the following results:

Step Training Loss Validation Loss WER
1000 2.505600 1.009531 0.614952
2000 1.186900 0.752440 0.422948
3000 0.947700 0.658266 0.358543
4000 0.817700 0.656034 0.356308
5000 0.741300 0.623420 0.314537
6000 0.694700 0.624534 0.294018
7000 0.653400 0.603341 0.286735
8000 0.616200 0.606606 0.285132
9000 0.594800 0.596215 0.277422
10000 0.590500 0.603380 0.274949

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: 30
  • lr_scheduler_type: linear

How to evaluate the model

from transformers import HubertForCTC, 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 = HubertForCTC.from_pretrained('TKU410410103/hubert-base-japanese-asr')
processor = Wav2Vec2Processor.from_pretrained("TKU410410103/hubert-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: 27.511982%
  • CER: 11.699897%

Framework versions

  • Transformers 4.39.1
  • Pytorch 2.2.1+cu118
  • Datasets 2.17.1