语言: eo
数据集:

• common_voice
  评估指标:
• wer
  标签:
• audio
• automatic-speech-recognition
• speech
• xlsr-fine-tuning-week
  许可证: apache-2.0
  模型索引:
• 名称: Wav2Vec2 Large 53 世界语 by Gunjan Chhablani
  结果:
  • 任务:
    名称: 语音识别
    类型: automatic-speech-recognition
    数据集:
    名称: Common Voice eo
    类型: common_voice
    参数: eo
    指标:
    • 名称: 测试 WER
      类型: wer
      值: 10.13

Wav2Vec2-Large-XLSR-53-世界语

基于facebook/wav2vec2-large-xlsr-53模型，使用Common Voice数据集对世界语进行微调。
使用此模型时，请确保语音输入采样率为16kHz。

使用方法

该模型可直接使用（无需语言模型），如下所示：

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

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

processor = Wav2Vec2Processor.from_pretrained('gchhablani/wav2vec2-large-xlsr-eo')  
model = Wav2Vec2ForCTC.from_pretrained('gchhablani/wav2vec2-large-xlsr-eo')  

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

# 数据预处理  
# 需要将音频文件读取为数组  
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("预测结果:", processor.batch_decode(predicted_ids))  
print("参考文本:", test_dataset["sentence"][:2])  

评估

该模型可按以下方式在Common Voice的世界语测试数据上进行评估。

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

import jiwer  

def chunked_wer(targets, predictions, chunk_size=None):  
    if chunk_size is None: return jiwer.wer(targets, predictions)  
    start = 0  
    end = chunk_size  
    H, S, D, I = 0, 0, 0, 0  
    while start < len(targets):  
        chunk_metrics = jiwer.compute_measures(targets[start:end], predictions[start:end])  
        H = H + chunk_metrics["hits"]  
        S = S + chunk_metrics["substitutions"]  
        D = D + chunk_metrics["deletions"]  
        I = I + chunk_metrics["insertions"]  
        start += chunk_size  
        end += chunk_size  
    return float(S + D + I) / float(H + S + D)  

test_dataset = load_dataset("common_voice", "eo", split="test") #TODO: 将{lang_id}替换为你的语言代码。确保代码是[此链接](https://huggingface.co/languages)中的*ISO代码*之一。  
wer = load_metric("wer")  

processor = Wav2Vec2Processor.from_pretrained('gchhablani/wav2vec2-large-xlsr-eo')  
model = Wav2Vec2ForCTC.from_pretrained('gchhablani/wav2vec2-large-xlsr-eo')  
model.to("cuda")  

chars_to_ignore_regex = """[\\\\\\\\,\\\\\\\\?\\\\\\\\.\\\\\\\\!\\\\\\\\-\\\\\\\\;\\\\\\\\:\\\\\\\\"\\\\\\\\“\\\\\\\\%\\\\\\\\‘\\\\\\\\”\\\\\\\\�\\\\\\\\„\\\\\\\\«\\\\\\\\(\\\\\\\\»\\\\\\\\)\\\\\\\\’\\\\\\\\']"""  
resampler = torchaudio.transforms.Resample(48_000, 16_000)  

# 数据预处理  
# 需要将音频文件读取为数组  
def speech_file_to_array_fn(batch):  
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace('—',' ').replace('–',' ')  
    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)  

# 数据预处理  
# 需要将音频文件读取为数组  
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 * chunked_wer(predictions=result["pred_strings"], targets=result["sentence"],chunk_size=5000)))  

测试结果: 10.13 %

训练

训练使用了Common Voice的train和validation数据集。代码可在此处找到。