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）

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 FunASR
pip3 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-zh

from funasr import AutoModel
from 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-streaming

from funasr import AutoModelchunk_size = [0, 10, 5] #[0, 10, 5] 600ms, [0, 8, 4] 480ms
encoder_chunk_look_back = 4 #number of chunks to lookback for encoder self-attention
decoder_chunk_look_back = 1 #number of encoder chunks to lookback for decoder cross-attentionmodel = AutoModel(model="./paraformer-zh-streaming", disable_update=True)import soundfile
import 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 - 1res = 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 sd
import numpy as np
import queue
from 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 = 1
CHUNK_MS = 600  # 每个 chunk 600ms
CHUNK_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")