基础模型: facebook/w2v-bert-2.0
库名称: transformers
语言:

• 乌克兰语
  许可证: "apache-2.0"
  任务类别:
• 自动语音识别
  标签:
• 音频
  数据集:
  • Yehor/openstt-uk
    评估指标:
  • 词错误率(WER)
    模型索引:
  • 名称: w2v-bert-uk-v2.1
    结果:
    • 任务:
      名称: 自动语音识别
      类型: automatic-speech-recognition
      数据集:
      名称: common_voice_10_0
      类型: common_voice_10_0
      配置: 乌克兰语
      拆分: 测试集
      参数: 乌克兰语
      指标:
      • 名称: 词错误率(WER)
        类型: wer
        值: 17.34
      • 名称: 字符错误率(CER)
        类型: cer
        值: 3.33

w2v-bert-uk v2.1版本

社区

• Discord社区: https://bit.ly/discord-uds
• 乌克兰语语音识别交流: https://t.me/speech_recognition_uk
• 乌克兰语语音合成交流: https://t.me/speech_synthesis_uk

其他乌克兰语模型请见: https://github.com/egorsmkv/speech-recognition-uk

概述

本模型是 https://huggingface.co/Yehor/w2v-bert-uk 的迭代版本

性能指标

• 声学模型(F16精度):
  • 词错误率(WER): 0.1734 (17.34%)
  • 字符错误率(CER): 0.0333 (3.33%)
  • 词语准确率: 82.66%
  • 字符准确率: 96.67%

在线演示

访问 https://huggingface.co/spaces/Yehor/w2v-bert-uk-v2.1-demo 体验模型对您上传音频的识别效果

使用方法

# 安装依赖: pip install -U torch soundfile transformers

import torch
import soundfile as sf
from transformers import AutoModelForCTC, Wav2Vec2BertProcessor

# 配置参数
model_name = 'Yehor/w2v-bert-uk-v2.1'
device = 'cuda:0' # 或cpu
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
sampling_rate = 16_000

# 加载模型
asr_model = AutoModelForCTC.from_pretrained(model_name, torch_dtype=torch_dtype).to(device)
processor = Wav2Vec2BertProcessor.from_pretrained(model_name)

paths = [
  '示例音频1.wav',
]

# 读取音频
audio_inputs = []
for path in paths:
  audio_input, _ = sf.read(path)
  audio_inputs.append(audio_input)

# 语音转写
inputs = processor(audio_inputs, sampling_rate=sampling_rate).input_features
features = torch.tensor(inputs).half().to(device)

with torch.inference_mode():
  logits = asr_model(features).logits

predicted_ids = torch.argmax(logits, dim=-1)
predictions = processor.batch_decode(predicted_ids)

# 输出结果
print('识别结果:')
print(predictions)

引用本工作

@misc {smoliakov_2025,
	author       = { {Smoliakov} },
	title        = { w2v-bert-uk-v2.1 (版本号 094c59d) },
	year         = 2025,
	url          = { https://huggingface.co/Yehor/w2v-bert-uk-v2.1 },
	doi          = { 10.57967/hf/4554 },
	publisher    = { Hugging Face }
}