Wav2vec2 Xlsr Khmer

🚀 Wav2Vec2-Large-XLSR-53-khmer

本項目是在高棉語數據集上對 facebook/wav2vec2-large-xlsr-53 模型進行微調得到的。使用了 Common Voice 和 OpenSLR Kh 數據集。使用該模型時，請確保輸入的語音採樣率為 16kHz。

模型信息

模型表現

🚀 快速開始

本模型可直接使用（無需語言模型），以下是使用示例。

💻 使用示例

基礎用法

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

!wget https://www.openslr.org/resources/42/km_kh_male.zip
!unzip km_kh_male.zip
!ls km_kh_male

colnames=['path','sentence'] 
df  = pd.read_csv('/content/km_kh_male/line_index.tsv',sep='\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\t',header=None,names = colnames)
df['path'] = '/content/km_kh_male/wavs/'+df['path'] +'.wav'

train, test = train_test_split(df, test_size=0.1)

test.to_csv('/content/km_kh_male/line_index_test.csv')

test_dataset = load_dataset('csv', data_files='/content/km_kh_male/line_index_test.csv',split = 'train')

processor = Wav2Vec2Processor.from_pretrained("gagan3012/wav2vec2-xlsr-nepali")
model = Wav2Vec2ForCTC.from_pretrained("gagan3012/wav2vec2-xlsr-nepali") 

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

# Preprocessing the datasets.
# We need to read the aduio files as arrays
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("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])

結果

預測結果：['पारानाको ब्राजिली राज्यमा रहेको राजधानी', 'देवराज जोशी त्रिभुवन विश्वविद्यालयबाट शिक्षाशास्त्रमा स्नातक हुनुहुन्छ'] 參考結果：['पारानाको ब्राजिली राज्यमा रहेको राजधानी', 'देवराज जोशी त्रिभुवन विश्वविद्यालयबाट शिक्षाशास्त्रमा स्नातक हुनुहुन्छ']

📚 詳細文檔

評估

可按以下方式在 Common Voice 的高棉語測試數據上對模型進行評估。

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
from sklearn.model_selection import train_test_split
import pandas as pd
from datasets import load_dataset

!wget https://www.openslr.org/resources/42/km_kh_male.zip
!unzip km_kh_male.zip
!ls km_kh_male

colnames=['path','sentence'] 
df  = pd.read_csv('/content/km_kh_male/line_index.tsv',sep='\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\t',header=None,names = colnames)
df['path'] = '/content/km_kh_male/wavs/'+df['path'] +'.wav'

train, test = train_test_split(df, test_size=0.1)

test.to_csv('/content/km_kh_male/line_index_test.csv')

test_dataset = load_dataset('csv', data_files='/content/km_kh_male/line_index_test.csv',split = 'train')
wer = load_metric("wer")
cer = load_metric("cer")


processor = Wav2Vec2Processor.from_pretrained("gagan3012/wav2vec2-xlsr-khmer")
model = Wav2Vec2ForCTC.from_pretrained("gagan3012/wav2vec2-xlsr-khmer") 
model.to("cuda")

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

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
    batch["text"] = re.sub(chars_to_ignore_regex, '', batch["text"]).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)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
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

cer = load_metric("cer")

result = test_dataset.map(evaluate, batched=True, batch_size=8)

print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["text"])))
print("CER: {:2f}".format(100 * cer.compute(predictions=result["pred_strings"], references=result["text"])))

測試結果：24.96 %
詞錯誤率（WER）：24.962519 字符錯誤率（CER）：6.950925

訓練

訓練腳本可在 此處 找到。

📄 許可證

本項目採用 Apache-2.0 許可證。