语言: el 数据集:

• common_voice 标签:
• 音频
• hf-asr-leaderboard
• 自动语音识别
• 语音
• xlsr-微调周 许可证: apache-2.0 模型索引:
• 名称: Lighteternal的XLSR Wav2Vec2希腊语模型 结果:
  • 任务: 名称: 语音识别 类型: automatic-speech-recognition 数据集: 名称: CommonVoice (EL), CSS10 (EL) 类型: CCS10 + mozilla-foundation/common_voice_7_0 参数: el 指标:
    • 名称: 测试WER 类型: wer 值: 10.497628
    • 名称: 测试CER 类型: cer 值: 2.875260

XLSR-Wav2Vec2自动语音识别(ASR)模型的希腊语(el)版本

由希腊军事学院和克里特技术大学开发

• 语言: el
• 许可证: apache-2.0
• 数据集: CommonVoice (EL), 364MB: https://commonvoice.mozilla.org/el/datasets + CSS10 (EL), 1.22GB: https://github.com/Kyubyong/css10
• 模型: XLSR-Wav2Vec2, 训练50个周期
• 指标: 词错误率(WER)

模型描述

更新: 我们使用来自CSS10的额外1.22GB数据集重复了微调过程。

Wav2Vec2是一个预训练的自动语音识别(ASR)模型，由Alexei Baevski、Michael Auli和Alex Conneau于2020年9月发布。在Wav2Vec2在英语ASR数据集LibriSpeech上展现出卓越性能后不久，Facebook AI提出了XLSR-Wav2Vec2。XLSR代表跨语言语音表示，指的是XLSR-Wav2Vec2能够学习跨多种语言有用的语音表示。

与Wav2Vec2类似，XLSR-Wav2Vec2从超过50种语言的数十万小时未标记语音中学习强大的语音表示。类似于BERT的掩码语言建模，该模型通过随机掩码特征向量，然后将它们传递给变压器网络，学习上下文相关的语音表示。

该模型在单个NVIDIA RTX 3080上训练了50个周期，大约8小时。

如何使用进行推理:

实时演示时，请确保语音文件采样率为16kHz。

测试CommonVoice提取的指令在ASR_Inference.ipynb中提供。代码片段如下:

#!/usr/bin/env python
# coding: utf-8

# 加载依赖项并定义预处理函数

from transformers import Wav2Vec2ForCTC
from transformers import Wav2Vec2Processor
from datasets import load_dataset, load_metric
import re
import torchaudio
import librosa
import numpy as np
from datasets import load_dataset, load_metric
import torch

chars_to_ignore_regex = '[\\\\\\\\,\\\\\\\\?\\\\\\\\.\\\\\\\\!\\\\\\\\-\\\\\\\\;\\\\\\\\:\\\\\\\\"\\\\\\\\“\\\\\\\\%\\\\\\\\‘\\\\\\\\”\\\\\\\\�]'

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

def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = speech_array[0].numpy()
    batch["sampling_rate"] = sampling_rate
    batch["target_text"] = batch["text"]
    return batch

def resample(batch):
    batch["speech"] = librosa.resample(np.asarray(batch["speech"]), 48_000, 16_000)
    batch["sampling_rate"] = 16_000
    return batch

def prepare_dataset(batch):
    # 检查所有文件是否具有正确的采样率
    assert (
        len(set(batch["sampling_rate"])) == 1
    ), f"确保所有输入的采样率均为{processor.feature_extractor.sampling_rate}."

    batch["input_values"] = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0]).input_values
    
    with processor.as_target_processor():
        batch["labels"] = processor(batch["target_text"]).input_ids
    return batch


# 加载模型和数据集处理器

model = Wav2Vec2ForCTC.from_pretrained("lighteternal/wav2vec2-large-xlsr-53-greek").to("cuda")
processor = Wav2Vec2Processor.from_pretrained("lighteternal/wav2vec2-large-xlsr-53-greek")


# 准备语音数据集以适合推理

common_voice_test = load_dataset("common_voice", "el", split="test")

common_voice_test = common_voice_test.remove_columns(["accent", "age", "client_id", "down_votes", "gender", "locale", "segment", "up_votes"])

common_voice_test = common_voice_test.map(remove_special_characters, remove_columns=["sentence"])

common_voice_test = common_voice_test.map(speech_file_to_array_fn, remove_columns=common_voice_test.column_names)

common_voice_test = common_voice_test.map(resample, num_proc=8)

common_voice_test = common_voice_test.map(prepare_dataset, remove_columns=common_voice_test.column_names, batch_size=8, num_proc=8, batched=True)


# 加载测试数据集

common_voice_test_transcription = load_dataset("common_voice", "el", split="test")


# 在随机样本上执行推理。更改"example"值以在不同CommonVoice提取上尝试推理

example = 123

input_dict = processor(common_voice_test["input_values"][example], return_tensors="pt", sampling_rate=16_000, padding=True)

logits = model(input_dict.input_values.to("cuda")).logits

pred_ids = torch.argmax(logits, dim=-1)

print("预测:")
print(processor.decode(pred_ids[0]))
# πού θέλεις να πάμε ρώτησε φοβισμένα ο βασιλιάς

print("\\\\
参考:")
print(common_voice_test_transcription["sentence"][example].lower())
# πού θέλεις να πάμε; ρώτησε φοβισμένα ο βασιλιάς.

评估

可以在Common Voice的希腊语测试数据上评估模型如下。

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

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

processor = Wav2Vec2Processor.from_pretrained("lighteternal/wav2vec2-large-xlsr-53-greek") 
model = Wav2Vec2ForCTC.from_pretrained("lighteternal/wav2vec2-large-xlsr-53-greek")
model.to("cuda")

chars_to_ignore_regex = '[\\\\\\\\,\\\\\\\\?\\\\\\\\.\\\\\\\\!\\\\\\\\-\\\\\\\\;\\\\\\\\:\\\\\\\\"\\\\\\\\“\\\\\\\\%\\\\\\\\‘\\\\\\\\”\\\\\\\\�]'
resampler = torchaudio.transforms.Resample(48_000, 16_000)

# 预处理数据集。
# 我们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()
  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)

# 预处理数据集。
# 我们需要将音频文件读取为数组

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: 10.497628 %

如何使用进行训练:

复制过程的指令和代码在Fine_Tune_XLSR_Wav2Vec2_on_Greek_ASR_with_🤗_Transformers.ipynb笔记本中提供。

指标