🚀 Wav2Vec2-Large-XLSR-53-粵語
本項目基於 Common Voice 語料庫 8.0 在粵語上對 facebook/wav2vec2-large-xlsr-53 進行了微調。使用該模型時,請確保輸入的語音採樣率為 16kHz。
模型信息
屬性 |
詳情 |
語言 |
粵語 |
數據集 |
Common Voice |
評估指標 |
字符錯誤率(CER) |
標籤 |
音頻、自動語音識別、語音、XLSR 微調周 |
許可證 |
Apache-2.0 |
模型表現
任務 |
數據集 |
評估指標 |
值 |
語音識別 |
Common Voice zh-HK |
測試字符錯誤率(Test CER) |
18.55% |
🚀 快速開始
本模型基於 Common Voice 語料庫 8.0 在粵語上對 facebook/wav2vec2-large-xlsr-53 進行了微調。使用該模型時,請確保輸入的語音採樣率為 16kHz。訓練使用了 Common Voice 中經過驗證的 train
和 dev
數據集。訓練腳本可在 https://github.com/holylovenia/wav2vec2-pretraining 找到。
💻 使用示例
基礎用法
import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
test_dataset = load_dataset("common_voice", "zh-HK", split="test[:2%]")
processor = Wav2Vec2Processor.from_pretrained("CAiRE/wav2vec2-large-xlsr-53-cantonese")
model = Wav2Vec2ForCTC.from_pretrained("CAiRE/wav2vec2-large-xlsr-53-cantonese")
def speech_file_to_array_fn(batch):
speech_array, sampling_rate = torchaudio.load(batch["path"])
resampler = torchaudio.transforms.Resample(sampling_rate, 16_000)
batch["speech"] = resampler(speech_array).squeeze().numpy()
return batch
test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset[:2]["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
with torch.no_grad():
logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset[:2]["sentence"])
高級用法
import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
test_dataset = load_dataset("common_voice", "zh-HK", split="test")
wer = load_metric("cer")
processor = Wav2Vec2Processor.from_pretrained("CAiRE/wav2vec2-large-xlsr-53-cantonese")
model = Wav2Vec2ForCTC.from_pretrained("CAiRE/wav2vec2-large-xlsr-53-cantonese")
model.to("cuda")
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\'\”\�]'
def speech_file_to_array_fn(batch):
batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
speech_array, sampling_rate = torchaudio.load(batch["path"])
resampler = torchaudio.transforms.Resample(sampling_rate, 16_000)
batch["speech"] = resampler(speech_array).squeeze().numpy()
return batch
test_dataset = test_dataset.map(speech_file_to_array_fn)
def evaluate(batch):
inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
with torch.no_grad():
logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
pred_ids = torch.argmax(logits, dim=-1)
batch["pred_strings"] = processor.batch_decode(pred_ids)
return batch
result = test_dataset.map(evaluate, batched=True, batch_size=8)
print("CER: {:2f}".format(100 * cer.compute(predictions=result["pred_strings"], references=result["sentence"])))
測試結果:字符錯誤率(CER):18.55 %
📄 引用
如果您使用了我們的代碼或模型,請引用以下文獻:
@inproceedings{lovenia2022ascend,
title={ASCEND: A Spontaneous Chinese-English Dataset for Code-switching in Multi-turn Conversation},
author={Lovenia, Holy and Cahyawijaya, Samuel and Winata, Genta Indra and Xu, Peng and Yan, Xu and Liu, Zihan and Frieske, Rita and Yu, Tiezheng and Dai, Wenliang and Barezi, Elham J and others},
booktitle={Proceedings of the 13th Language Resources and Evaluation Conference (LREC)},
year={2022}
}