Wav2vec2 Large Xlsr Cantonese

🚀 Wav2Vec2-Large-XLSR-53-粵語

本模型基於Common Voice粵語數據，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

模型表現

🚀 快速開始

✨ 主要特性

• 基於預訓練模型facebook/wav2vec2-large-xlsr-53微調，在粵語語音識別任務上表現良好。
• 可直接使用，無需語言模型。

📦 安裝指南

文檔未提及安裝步驟，可參考Hugging Face相關庫的安裝方法，如：

pip install transformers datasets torchaudio jiwer

💻 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("ctl/wav2vec2-large-xlsr-cantonese") 
model = Wav2Vec2ForCTC.from_pretrained("ctl/wav2vec2-large-xlsr-cantonese")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
	speech_array, sampling_rate = torchaudio.load(batch["path"])
	batch["speech"] = resampler(speech_array).squeeze().numpy()
	return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset["speech"][:2], 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["sentence"][:2])

📚 詳細文檔

評估

該模型可在Common Voice的中文（香港）測試數據上進行評估，示例代碼如下：

!pip install jiwer
import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
import argparse

lang_id = "zh-HK" 
model_id = "ctl/wav2vec2-large-xlsr-cantonese"

chars_to_ignore_regex = '[\,\?\.\!\-\;\:"\“\%\‘\”\�\．\⋯\！\－\：\–\。\》\,\）\,\？\；\～\~\…\︰\，\（\」\‧\《\﹔\、\—\／\,\「\﹖\·\']'

test_dataset = load_dataset("common_voice", f"{lang_id}", split="test") 
cer = load_metric("cer")

processor = Wav2Vec2Processor.from_pretrained(f"{model_id}") 
model = Wav2Vec2ForCTC.from_pretrained(f"{model_id}") 
model.to("cuda")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
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"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
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=16)

print("CER: {:2f}".format(100 * cer.compute(predictions=result["pred_strings"], references=result["sentence"])))

測試結果：15.51 %

訓練

訓練使用了Common Voice的train和validation數據集。訓練腳本將發佈在此處。

📄 許可證

本項目採用Apache-2.0許可證。