基础模型:

• antony66/whisper-large-v3-russian
• bond005/whisper-large-v3-ru-podlodka 语言:
• 俄语 库名称: transformers 标签:
• 自动语音识别
• whisper
• 俄语
• 模型合并工具
• 合并 数据集:
• mozilla-foundation/common_voice_17_0
• bond005/taiga_speech_v2
• bond005/podlodka_speech
• bond005/rulibrispeech 评估指标:
• 词错误率

新版已发布: Apel-sin/whisper-large-v3-russian-ties-podlodka-v1.2

模型详情

本模型采用TIES合并方法融合而成。

合并方法: ties
参数:
  连接密度: 0.85
  编码器权重:
    - 0.65
    - 0.35
  解码器权重:
    - 0.6
    - 0.4
模型:
  模型A: "/mnt/cloud/llm/whisper/whisper-large-v3-russian"
  模型B: "/mnt/cloud/llm/whisper/whisper-large-v3-ru-podlodka"
输出目录: "/mnt/cloud/llm/whisper/whisper-large-v3-russian-ties-podlodka"

简易API服务

可配合开源OpenAI兼容API服务器使用: https://github.com/kreolsky/whisper-api-server/

使用指南

处理电话录音时强烈建议先进行音量预处理，例如:

sox 录音.wav -r 8000 标准化录音.wav norm -0.5 compand 0.3,1 -90,-90,-70,-50,-40,-15,0,0 -7 0 0.15

ASR代码参考示例:

import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor, pipeline

torch_dtype = torch.bfloat16 # 设置首选数据类型

device = 'cpu'
if torch.cuda.is_available():
    device = 'cuda'
elif torch.backends.mps.is_available():
    device = 'mps'
    setattr(torch.distributed, "is_initialized", lambda : False) # 临时补丁
device = torch.device(device)

whisper = WhisperForConditionalGeneration.from_pretrained(
    "antony66/whisper-large-v3-russian", torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True,
    # 若GPU支持可添加 attn_implementation="flash_attention_2"
)

processor = WhisperProcessor.from_pretrained("antony66/whisper-large-v3-russian")

asr_pipeline = pipeline(
    "automatic-speech-recognition",
    model=whisper,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=256,
    chunk_length_s=30,
    batch_size=16,
    return_timestamps=True,
    torch_dtype=torch_dtype,
    device=device,
)

# 读取音频文件示例:
from io import BytesIO
wav = BytesIO()
with open('标准化录音.wav', 'rb') as f:
    wav.write(f.read())
wav.seek(0)

# 获取转写结果
asr = asr_pipeline(wav, generate_kwargs={"language": "russian", "max_new_tokens": 256}, return_timestamps=False)

print(asr['text'])

开发进展

当前为开发测试版本，目标是优化电话语音识别效果。如有优质数据集或改进建议，欢迎联系贡献，我们将不胜感激。