语言: 罗马尼亚语
数据集:

• common_voice
  评估指标:
• wer (词错误率)
  标签:
• 音频
• 自动语音识别
• 语音
• xlsr微调周
  许可证: apache-2.0
  模型索引:
• 名称: Anton Lozhkov的罗马尼亚语XLSR Wav2Vec2 Large 53
  结果:
  • 任务:
    名称: 语音识别
    类型: automatic-speech-recognition
    数据集:
    名称: Common Voice 罗马尼亚语
    类型: common_voice
    参数: ro
    指标:
    • 名称: 测试WER
      类型: wer
      值: 24.84

Wav2Vec2-Large-XLSR-53-罗马尼亚语版

本模型基于facebook/wav2vec2-large-xlsr-53，使用Common Voice的罗马尼亚语数据集进行微调。
使用时请确保语音输入采样率为16kHz。

使用方法

可直接使用该模型（无需语言模型），示例如下：

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

test_dataset = load_dataset("common_voice", "ro", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("anton-l/wav2vec2-large-xlsr-53-romanian")
model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-romanian")

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

# 数据预处理
# 需要将音频文件读取为数组
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("预测结果:", processor.batch_decode(predicted_ids))
print("参考文本:", test_dataset["sentence"][:2])

评估

可通过以下方式在Common Voice罗马尼亚语测试集上评估模型：

import torch
import torchaudio
import urllib.request
import tarfile
import pandas as pd
from tqdm.auto import tqdm
from datasets import load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

# 为节省磁盘空间，直接下载原始数据而非使用HF数据集
data_url = "https://voice-prod-bundler-ee1969a6ce8178826482b88e843c335139bd3fb4.s3.amazonaws.com/cv-corpus-6.1-2020-12-11/ro.tar.gz"
filestream = urllib.request.urlopen(data_url)
data_file = tarfile.open(fileobj=filestream, mode="r|gz")
data_file.extractall()

wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("anton-l/wav2vec2-large-xlsr-53-romanian")
model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-romanian")
model.to("cuda")

cv_test = pd.read_csv("cv-corpus-6.1-2020-12-11/ro/test.tsv", sep='\t')
clips_path = "cv-corpus-6.1-2020-12-11/ro/clips/"

def clean_sentence(sent):
    sent = sent.lower()
    # 将非字母字符替换为空格
    sent = "".join(ch if ch.isalpha() else " " for ch in sent)
    # 去除重复空格
    sent = " ".join(sent.split())
    return sent

targets = []
preds = []

for i, row in tqdm(cv_test.iterrows(), total=cv_test.shape[0]):
    row["sentence"] = clean_sentence(row["sentence"])
    speech_array, sampling_rate = torchaudio.load(clips_path + row["path"])
    resampler = torchaudio.transforms.Resample(sampling_rate, 16_000)
    row["speech"] = resampler(speech_array).squeeze().numpy()

    inputs = processor(row["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)

    targets.append(row["sentence"])
    preds.append(processor.batch_decode(pred_ids)[0])

print("WER: {:2f}".format(100 * wer.compute(predictions=preds, references=targets)))

测试结果: 24.84%

训练过程

使用Common Voice的train和validation数据集进行训练。