Wav2Vec2-Large-XLSR-53-Marathi

使用 Open SLR64 数据集在马拉地语上进行了 facebook/wav2vec2-large-xlsr-53 的微调。使用该模型时，请确保语音输入采样率为16kHz。此数据仅包含女性声音，但该模型也适用于男性声音。在Google Colab Pro的Tesla P100 16GB GPU上训练。测试集的词错误率（WER）为12.70%。

使用方法

模型可以直接使用，无需语言模型。假设你的数据集有马拉地语的actual_text和path_in_folder列:

import torch, torchaudio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

#Since marathi is not present on Common Voice, script for reading the below dataset can be picked up from the eval script below
mr_test_dataset = all_data['test']

processor = Wav2Vec2Processor.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi") 
model = Wav2Vec2ForCTC.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi") 

resampler = torchaudio.transforms.Resample(48_000, 16_000) #first arg - input sample, second arg - output sample
# 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_in_folder"])
  batch["speech"] = resampler(speech_array).squeeze().numpy()
  return batch
mr_test_dataset = mr_test_dataset.map(speech_file_to_array_fn)
inputs = processor(mr_test_dataset["speech"][:5], 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:", mr_test_dataset["actual_text"][:5])

评估

在Open SLR-64的马拉地语数据的10%上进行了评估。

import os, re, torch, torchaudio
from datasets import Dataset, load_metric
import pandas as pd
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

#below is a custom script to be used for reading marathi dataset since its not present on the Common Voice
dataset_path = "./OpenSLR-64_Marathi/mr_in_female/" #TODO : include the path of the dataset extracted from http://openslr.org/64/
audio_df = pd.read_csv(os.path.join(dataset_path,'line_index.tsv'),sep='\t',header=None)
audio_df.columns = ['path_in_folder','actual_text']
audio_df['path_in_folder'] = audio_df['path_in_folder'].apply(lambda x: dataset_path + x + '.wav')
audio_df = audio_df.sample(frac=1, random_state=2020).reset_index(drop=True) #seed number is important for reproducibility of WER score
all_data = Dataset.from_pandas(audio_df)
all_data = all_data.train_test_split(test_size=0.10,seed=2020) #seed number is important for reproducibility of WER score

mr_test_dataset = all_data['test']
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi")
model = Wav2Vec2ForCTC.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi") 
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["actual_text"] = re.sub(chars_to_ignore_regex, '', batch["actual_text"]).lower()
  speech_array, sampling_rate = torchaudio.load(batch["path_in_folder"])
  batch["speech"] = resampler(speech_array).squeeze().numpy()
  return batch
mr_test_dataset = mr_test_dataset.map(speech_file_to_array_fn)
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 = mr_test_dataset.map(evaluate, batched=True, batch_size=8)
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["actual_text"])))

训练

训练-测试比例为90:10。训练笔记本的Colab链接 here 。

训练配置和总结

权重和偏差的运行总结 here