语言: is
数据集:

• malromur
  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:
• 示例标题: Malromur 样本 1608
  来源: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/resolve/main/sample1608.flac
• 示例标题: Malromur 样本 3860
  来源: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/resolve/main/sample3860.flac
  模型索引:
• 名称: Mehrdad Farahani 的 XLSR Wav2Vec2 冰岛语模型
  结果:
  • 任务:
    名称: 语音识别
    类型: automatic-speech-recognition
    数据集:
    名称: Malromur is
    类型: malromur
    参数: lt
    指标:
    • 名称: 测试 WER
      类型: wer
      值: 09.21

Wav2Vec2-Large-XLSR-53-冰岛语

使用 Malromur 对 facebook/wav2vec2-large-xlsr-53 进行冰岛语微调。使用此模型时，请确保语音输入采样率为16kHz。

使用方式

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
!pip install num2words

标准化器

# num2word 包
# 原始来源: https://github.com/savoirfairelinux/num2words
!mkdir -p ./num2words
!wget -O num2words/__init__.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/__init__.py
!wget -O num2words/base.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/base.py
!wget -O num2words/compat.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/compat.py
!wget -O num2words/currency.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/currency.py
!wget -O num2words/lang_EU.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/lang_EU.py
!wget -O num2words/lang_IS.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/lang_IS.py
!wget -O num2words/utils.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/num2words/utils.py

# 基于性别和年龄选择的 Malromur_test
!wget -O malromur_test.csv https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/malromur_test.csv

# 标准化器
!wget -O normalizer.py https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-icelandic/raw/main/normalizer.py

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

from normalizer import Normalizer

normalizer = Normalizer(lang="is")


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)
    return batch


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

dataset = load_dataset("csv", data_files={"test": "./malromur_test.csv"})["test"]
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"do_lastspace_removing": True, "text_key_name": "cleaned_sentence"},
    remove_columns=list(set(dataset.column_names) - set(['cleaned_sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict, batched=True, batch_size=8)

max_items = np.random.randint(0, len(result), 20).tolist()
for i in max_items:
    reference, predicted =  result["cleaned_sentence"][i], result["predicted"][i]
    print("参考:", reference)
    print("预测:", predicted)
    print('---')

输出:

参考: eða eitthvað annað dýr
预测: eða eitthvað annað dýr
---
参考: oddgerður
预测: oddgerður
---
参考: eiðný
预测: eiðný
---
参考: löndum
预测: löndum
---
参考: tileinkaði bróður sínum markið
预测: tileinkaði bróður sínum markið
---
参考: þetta er svo mikill hégómi
预测: þetta er svo mikill hégómi
---
参考: timarit is
预测: timarit is
---
参考: stefna strax upp aftur
预测: stefna strax upp aftur
---
参考: brekkuflöt
预测: brekkuflöt
---
参考: áætlunarferð frestað vegna veðurs
预测: áætluna ferð frestað vegna veðurs
---
参考: sagði af sér vegna kláms
预测: sagði af sér vegni kláms
---
参考: grímúlfur
预测: grímúlgur
---
参考: lýsti sig saklausan
预测: lýsti sig saklausan
---
参考: belgingur is
预测: belgingur is
---
参考: sambía
预测: sambía
---
参考: geirastöðum
预测: geirastöðum
---
参考: varð tvisvar fyrir eigin bíl
预测: var tvisvar fyrir eigin bíl
---
参考: reykjavöllum
预测: reykjavöllum
---
参考: miklir menn eru þeir þremenningar
预测: miklir menn eru þeir þremenningar
---
参考: handverkoghonnun is
预测: handverkoghonnun is
---

评估

可以在 Malromur 的测试数据上评估模型如下：

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

from normalizer import Normalizer

normalizer = Normalizer(lang="is")


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)
    return batch


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

dataset = load_dataset("csv", data_files={"test": "./malromur_test.csv"})["test"]
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"do_lastspace_removing": True, "text_key_name": "cleaned_sentence"},
    remove_columns=list(set(dataset.column_names) - set(['cleaned_sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict, batched=True, batch_size=8)

wer = load_metric("wer")

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

测试结果:

• WER: 09.21%

训练与报告

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

训练状态可在此处查看 此处

训练脚本可在此处找到 此处

问题？

请在 Wav2Vec 仓库提交一个 Github issue。