Wav2Vec2-Large-XLSR-53-Thai

在泰语中使用 Common Voice 的Fine-tuned facebook/wav2vec2-large-xlsr-53 。使用此模型时，请确保语音输入采样率为16kHz。

用法

可以直接使用该模型（无需语言模型），方法如下：

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from pythainlp.tokenize import word_tokenize

test_dataset = load_dataset("common_voice", "th", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("sakares/wav2vec2-large-xlsr-thai-demo")
model = Wav2Vec2ForCTC.from_pretrained("sakares/wav2vec2-large-xlsr-thai-demo")

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

## For Thai NLP Library, please feel free to check https://pythainlp.github.io/docs/2.2/api/tokenize.html
def th_tokenize(batch):
    batch["sentence"] = " ".join(word_tokenize(batch["sentence"], engine="newmm"))
    return batch

# 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).map(th_tokenize)
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])

用法脚本 here

评估

可以按如下方式评估Common Voice的{language}测试数据：

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from pythainlp.tokenize import word_tokenize
import re

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

processor = Wav2Vec2Processor.from_pretrained("sakares/wav2vec2-large-xlsr-thai-demo")
model = Wav2Vec2ForCTC.from_pretrained("sakares/wav2vec2-large-xlsr-thai-demo")
model.to("cuda")

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

## For Thai NLP Library, please feel free to check https://pythainlp.github.io/docs/2.2/api/tokenize.html
def th_tokenize(batch):
    batch["sentence"] = " ".join(word_tokenize(batch["sentence"], engine="newmm"))
    return batch

# Preprocessing the datasets.
# We need to read the aduio files as arrays
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).map(th_tokenize)

# 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

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：44.46%评估脚本 here

训练

使用了Common Voice的训练数据集和验证数据集进行训练。

训练所使用的脚本可以在 here 处找到。