语言: br 数据集:

• common_voice 指标:
• wer 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: XLSR Wav2Vec2 Breton by Cahya 结果:
  • 任务: 名称: 语音识别 类型: automatic-speech-recognition 数据集: 名称: Common Voice br 类型: common_voice 参数: br 指标:
    • 名称: 测试WER 类型: wer 值: 41.71

Wav2Vec2-Large-XLSR-Breton

基于facebook/wav2vec2-large-xlsr-53在布列塔尼语Common Voice数据集上微调。使用该模型时，请确保语音输入采样率为16kHz。

使用方法

该模型可直接使用（无需语言模型）如下：

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

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

processor = Wav2Vec2Processor.from_pretrained("cahya/wav2vec2-large-xlsr-breton")
model = Wav2Vec2ForCTC.from_pretrained("cahya/wav2vec2-large-xlsr-breton")

chars_to_ignore_regex = '[\\,\,\?\.\!\;\:\"\“\%\”\�\(\)\/\«\»\½\…]'

# 数据预处理
# 需要将音频文件读取为数组
def speech_file_to_array_fn(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower() + " "
    batch["sentence"] = batch["sentence"].replace("ʼ", "'")
    batch["sentence"] = batch["sentence"].replace("’", "'")
    batch["sentence"] = batch["sentence"].replace('‘', "'")
    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("预测结果:", processor.batch_decode(predicted_ids))
print("参考文本:", test_dataset[:2]["sentence"])

上述代码对前两个样本的预测结果如下：

预测结果: ["ne' ler ket don a-benn us netra pa vez zer nic'hed evel-si", 'an eil hag egile']
参考文本: ['"n\'haller ket dont a-benn eus netra pa vezer nec\'het evel-se." ', 'an eil hag egile. ']

评估

该模型可在Common Voice的布列塔尼语测试数据上如下评估：

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

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

processor = Wav2Vec2Processor.from_pretrained("cahya/wav2vec2-large-xlsr-breton")
model = Wav2Vec2ForCTC.from_pretrained("cahya/wav2vec2-large-xlsr-breton") 
model.to("cuda")

chars_to_ignore_regex = '[\\,\,\?\.\!\;\:\"\“\%\”\�\(\)\/\«\»\½\…]'

# 数据预处理
# 需要将音频文件读取为数组
def speech_file_to_array_fn(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower() + " "
    batch["sentence"] = batch["sentence"].replace("ʼ", "'")
    batch["sentence"] = batch["sentence"].replace("’", "'")
    batch["sentence"] = batch["sentence"].replace('‘', "'")
    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)

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

测试结果: 41.71 %

训练

使用了Common Voice的train、validation等数据集进行训练（待补充细节）

训练脚本可在此处找到（即将上线）