语言: 爱沙尼亚语
数据集:

• common_voice
  标签:
• 音频
• 自动语音识别
• 语音
• xlsr微调周
  许可证: apache-2.0
  小部件:
• 标签: Common Voice样本1123
  源: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-estonian/resolve/main/sample1123.flac
• 标签: Common Voice样本910
  源: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-estonian/resolve/main/sample910.flac
  模型索引:
• 名称: Mehrdad Farahani的爱沙尼亚语XLSR Wav2Vec2
  结果:
  • 任务:
    名称: 语音识别
    类型: automatic-speech-recognition
    数据集:
    名称: Common Voice et
    类型: common_voice
    参数: et
    指标:
    • 名称: 测试WER
      类型: wer
      值: 33.93

Wav2Vec2-Large-XLSR-53-爱沙尼亚语

使用Common Voice对facebook/wav2vec2-large-xlsr-53进行爱沙尼亚语微调。使用此模型时，请确保语音输入采样率为16kHz。

使用方法

该模型可以直接使用（无需语言模型）如下：

要求

# 所需包
!pip install git+https://github.com/huggingface/datasets.git
!pip install git+https://github.com/huggingface/transformers.git
!pip install torchaudio
!pip install librosa
!pip install jiwer

预测

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

import numpy as np
import re
import string

import IPython.display as ipd

chars_to_ignore = [
    ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�",
    "#", "!", "?", "«", "»", "(", ")", "؛", ",", "?", ".", "!", "-", ";", ":", '"', 
    "“", "%", "‘", "�", "–", "…", "_", "”", '“', '„'
]
chars_to_mapping = {
"\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ",
}

def multiple_replace(text, chars_to_mapping):
    pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
    return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))

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

def normalizer(batch, chars_to_ignore, chars_to_mapping):
    chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]"""
    text = batch["sentence"].lower().strip()
    
    text = text.replace("\u0307", " ").strip()
    text = multiple_replace(text, chars_to_mapping)
    text = remove_special_characters(text, chars_to_ignore_regex)

    batch["sentence"] = text
    return batch


def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array = speech_array.squeeze().numpy()
    speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000)

    batch["speech"] = speech_array
    return batch


def predict(batch):
    features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    input_values = features.input_values.to(device)
    attention_mask = features.attention_mask.to(device)

    with torch.no_grad():
        logits = model(input_values, attention_mask=attention_mask).logits 
        
    pred_ids = torch.argmax(logits, dim=-1)

    batch["predicted"] = processor.batch_decode(pred_ids)[0]
    return batch


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-estonian")
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-estonian").to(device)

dataset = load_dataset("common_voice", "et", split="test[:1%]")
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping},
    remove_columns=list(set(dataset.column_names) - set(['sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict)

max_items = np.random.randint(0, len(result), 10).tolist()
for i in max_items:
    reference, predicted =  result["sentence"][i], result["predicted"][i]
    print("参考:", reference)
    print("预测:", predicted)
    print('---')

输出:

参考: õhulossid lagunevad ning ees ootab maapind 
预测: õhulassid lagunevad ning ees ootab maapind
---
参考: milliseks kiievisse pääsemise nimel võistlev muusik soime muusikamaastiku hetkeseisu hindab ning kas ta ka ennast sellel tulevikus tegutsemas näeb kuuled videost 
预测: milliseks gievisse pääsemise nimel võitlev muusiks soime muusikama aastiku hetke seisu hindab ning kas ta ennast selle tulevikus tegutsemast näeb kuulad videost
---
参考: näiteks kui pool seina on tehtud tekib tunne et tahaks tegelikult natuke teistsugust ja hakkame otsast peale 
预测: näiteks kui pool seine on tehtud tekib tunnetahaks tegelikult matuka teistsugust jahappanna otsast peane
---
参考: neuroesteetilised katsed näitavad et just nägude vaatlemine aktiveerib inimese aju esteetilist keskust 
预测: neuroaisteetiliselt katsed näitaval et just nägude vaatlemine aptiveerid inimese aju est eedilist keskust
---
参考: paljud inimesed kindlasti kadestavad teid kuid ei julge samamoodi vabalt võtta 
预测: paljud inimesed kindlasti kadestavadteid kuid ei julge sama moodi vabalt võtta
---
参考: parem on otsida pileteid inkognito veebi kaudu 
预测: parem on otsida pileteid ning kognitu veebikaudu
---
参考: ja vot siin ma jäin vaikseks 
预测: ja vat siisma ja invaikseks
---
参考: mida sa iseendale juubeli puhul soovid 
预测: mida saise endale jubeli puhul soovid
---
参考: kuumuse ja kõrge temperatuuri tõttu kuivas tühjadel karjamaadel rohi mis muutus kergesti süttivaks 
预测: kuumuse ja kõrge temperatuuri tõttu kuivast ühjadal karjamaadel rohi mis muutus kergesti süttivaks
---
参考: ilmselt on inimesi kelle jaoks on see hea lahendus 
预测: ilmselt on inimesi kelle jaoks on see hea lahendus
---

评估

可以在Common Voice的爱沙尼亚语测试数据上评估该模型如下。

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

import numpy as np
import re
import string


chars_to_ignore = [
    ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�",
    "#", "!", "?", "«", "»", "(", ")", "؛", ",", "?", ".", "!", "-", ";", ":", '"', 
    "“", "%", "‘", "�", "–", "…", "_", "”", '“', '„'
]
chars_to_mapping = {
    "\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ",
}

def multiple_replace(text, chars_to_mapping):
    pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
    return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))

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

def normalizer(batch, chars_to_ignore, chars_to_mapping):
    chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]"""
    text = batch["sentence"].lower().strip()
    
    text = text.replace("\u0307", " ").strip()
    text = multiple_replace(text, chars_to_mapping)
    text = remove_special_characters(text, chars_to_ignore_regex)

    batch["sentence"] = text
    return batch


def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array = speech_array.squeeze().numpy()
    speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000)

    batch["speech"] = speech_array
    return batch


def predict(batch):
    features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    input_values = features.input_values.to(device)
    attention_mask = features.attention_mask.to(device)

    with torch.no_grad():
        logits = model(input_values, attention_mask=attention_mask).logits 
        
    pred_ids = torch.argmax(logits, dim=-1)

    batch["predicted"] = processor.batch_decode(pred_ids)[0]
    return batch


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-estonian")
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-estonian").to(device)

dataset = load_dataset("common_voice", "et", split="test")
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping},
    remove_columns=list(set(dataset.column_names) - set(['sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict)

wer = load_metric("wer")

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

测试结果:

• WER: 33.93%

训练与报告

训练使用了Common Voice的train和validation数据集。

您可以在此处查看训练状态here

训练脚本可在此处找到here