阿里 FunASR 开源中文语音识别大模型应用示例（准确率比faster-whisper高）_funasr官网
文章目录
- Github
- 官网
- 简介
- 模型
- 安装
- 非流式应用示例
- 流式应用示例
Github
- https://github.com/modelscope/FunASR
官网
- https://www.funasr.com/#/
简介
FunASR是一个基础语音识别工具包，提供多种功能，包括语音识别（ASR）、语音端点检测（VAD）、标点恢复、语言模型、说话人验证、说话人分离和多人对话语音识别等。FunASR提供了便捷的脚本和教程，支持预训练好的模型的推理与微调。
我们在ModelScope与huggingface上发布了大量开源数据集或者海量工业数据训练的模型，可以通过我们的模型仓库了解模型的详细信息。代表性的Paraformer非自回归端到端语音识别模型具有高精度、高效率、便捷部署的优点，支持快速构建语音识别服务，详细信息可以阅读(服务部署文档)。
模型
- https://huggingface.co/funasr
安装
pip3 install -U funasr# 或从源码安装git clone https://github.com/alibaba/FunASR.git && cd FunASRpip3 install -e ./非流式应用示例
- https://huggingface.co/funasr/paraformer-zh
- 下载模型
# 语音识别，带时间戳，非流式git clone https://huggingface.co/funasr/paraformer-zhfrom funasr import AutoModelfrom funasr.utils.postprocess_utils import rich_transcription_postprocessmodel = AutoModel( model=\'./paraformer-zh\', vad_model=\"fsmn-vad\", vad_kwargs={\"max_single_segment_time\": 30000}, device=\"cpu\", disable_update=True, disable_pbar=True)res = model.generate( input=f\"./zh_20250329_105237.wav\", cache={}, language=\"auto\", # \"zn\", \"en\", \"yue\", \"ja\", \"ko\", \"nospeech\" use_itn=True, batch_size_s=60, merge_vad=True, # merge_length_s=15,)text = rich_transcription_postprocess(res[0][\"text\"])print(text)流式应用示例
- https://huggingface.co/funasr/paraformer-zh-streaming
- 下载模型
# 语音识别、流媒体git clone https://huggingface.co/funasr/paraformer-zh-streamingfrom funasr import AutoModelchunk_size = [0, 10, 5] #[0, 10, 5] 600ms, [0, 8, 4] 480msencoder_chunk_look_back = 4 #number of chunks to lookback for encoder self-attentiondecoder_chunk_look_back = 1 #number of encoder chunks to lookback for decoder cross-attentionmodel = AutoModel(model=\"./paraformer-zh-streaming\", disable_update=True)import soundfileimport osspeech, sample_rate = soundfile.read(\'./zh_20250329_105237.wav\')chunk_stride = chunk_size[1] * 960 # 600mscache = {}# 计算总共需要处理的音频块数total_chunk_num = int(len((speech)-1)/chunk_stride+1)# 遍历所有音频块，逐块识别for i in range(total_chunk_num): speech_chunk = speech[i*chunk_stride:(i+1)*chunk_stride] is_final = i == total_chunk_num - 1 res = model.generate(input=speech_chunk, cache=cache, is_final=is_final, chunk_size=chunk_size, encoder_chunk_look_back=encoder_chunk_look_back, decoder_chunk_look_back=decoder_chunk_look_back) print(res)注：chunk_size为流式传输延迟的配置。 [0,10,5]表示实时显示粒度为1060=600ms，前瞻信息为560=300ms。每次推理的输入为600ms（样本点为16000*0.6=960），输出为对应的文本。对于最后一段语音的输入，is_final=True需要设置 ，强制输出最后一个单词。
import sounddevice as sdimport numpy as npimport queuefrom funasr import AutoModelmodel = AutoModel( model=\"./paraformer-zh-streaming\", disable_update=True, disable_pbar=True )# 录音参数SAMPLE_RATE = 16000 # FunASR 需要 16kHz 采样率CHANNELS = 1CHUNK_MS = 600 # 每个 chunk 600msCHUNK_SIZE = CHUNK_MS * 16 # 600ms 音频 = 9600 采样点LOOK_BACK = 4 # Encoder 关注的 chunk 数量# 音频缓存audio_queue = queue.Queue()cache = {} # 识别缓存# 录音回调def callback(indata, frames, time, status): if status: print(f\"音频流错误: {status}\") audio_queue.put(indata.copy()) # 存入队列# 开始监听麦克风print(\"正在监听麦克风... (Ctrl+C 退出)\")try: with sd.InputStream(samplerate=SAMPLE_RATE, channels=CHANNELS, callback=callback, blocksize=CHUNK_SIZE): buffer = np.array([], dtype=np.float32) # 音频缓冲区 while True: if not audio_queue.empty(): # 获取音频数据 audio_chunk = audio_queue.get().flatten() buffer = np.append(buffer, audio_chunk) # 添加到缓冲区 # 每 600ms 进行一次识别 while len(buffer) >= CHUNK_SIZE: chunk = buffer[:CHUNK_SIZE] buffer = buffer[CHUNK_SIZE:] # 移除已处理部分 # 执行语音识别 res = model.generate( input=chunk, cache=cache, is_final=False, chunk_size=[0, 10, 5], encoder_chunk_look_back=LOOK_BACK, decoder_chunk_look_back=1 ) print(\"识别结果:\", res[0][\'text\'])except KeyboardInterrupt: print(\"停止监听\")
