Wav2vec2 Large Xlsr 53 Eu

🚀 Wav2Vec2-Large-XLSR-53-EU

本項目是在巴斯克語上對 facebook/wav2vec2-large-xlsr-53 模型進行微調的成果，使用的數據集為 Common Voice。使用該模型時，請確保輸入的語音採樣率為 16kHz。

🚀 快速開始

本模型可直接使用（無需語言模型），具體操作如下。

💻 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", "eu", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("pcuenq/wav2vec2-large-xlsr-53-eu")
model = Wav2Vec2ForCTC.from_pretrained("pcuenq/wav2vec2-large-xlsr-53-eu")

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

# Preprocessing the datasets.
# We need to read the audio 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])

🔧 技術細節

評估

該模型可以在 Common Voice 的巴斯克語測試數據上進行如下評估：

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re

test_dataset = load_dataset("common_voice", "eu", split="test")
wer = load_metric("wer")

model_name = "pcuenq/wav2vec2-large-xlsr-53-eu"

processor = Wav2Vec2Processor.from_pretrained(model_name)
model = Wav2Vec2ForCTC.from_pretrained(model_name)
model.to("cuda")

## Text pre-processing

chars_to_ignore_regex = '[\,\¿\?\.\¡\!\-\;\:\"\“\%\‘\”\\…\’\ː\'\‹\›\`\´\®\—\→]'
chars_to_ignore_pattern = re.compile(chars_to_ignore_regex)

def remove_special_characters(batch):
    batch["sentence"] = chars_to_ignore_pattern.sub('', batch["sentence"]).lower() + " "
    return batch

## Audio pre-processing

import librosa
def speech_file_to_array_fn(batch):
    speech_array, sample_rate = torchaudio.load(batch["path"])
    batch["speech"] = librosa.resample(speech_array.squeeze().numpy(), sample_rate, 16_000)
    return batch

# Text transformation and audio resampling
def cv_prepare(batch):
    batch = remove_special_characters(batch)
    batch = speech_file_to_array_fn(batch)
    return batch

# Number of CPUs or None
num_proc = 16
test_dataset = test_dataset.map(cv_prepare, remove_columns=['path'], num_proc=num_proc)

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)

# WER Metric computation
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))

測試結果：15.34 %

訓練

訓練使用了 Common Voice 的 train 和 validation 數據集。訓練共進行了 22 + 20 個 epoch，使用了以下參數：

• 批量大小：16，梯度累積步數：2。
• 學習率：2.5e-4
• 激活丟棄率：0.05
• 注意力丟棄率：0.1
• 隱藏層丟棄率：0.05
• 特徵投影丟棄率：0.05
• 掩碼時間概率：0.08
• 層丟棄率：0.05

📄 許可證

本項目採用 Apache-2.0 許可證。