Merge pull request #287 from chmod740/add_fun_asr

feat:添加funasr语音识别框架支持

requirements.txt

@@ -19,4 +19,5 @@ openai
 apscheduler
 asyncio
 edge-tts
-nest_asyncio
+nest_asyncio
+funasr_onnx

robot/ASR.py

@@ -1,7 +1,7 @@
 # -*- coding: utf-8 -*-
 import json
 from aip import AipSpeech
-from .sdk import TencentSpeech, AliSpeech, XunfeiSpeech, BaiduSpeech
+from .sdk import TencentSpeech, AliSpeech, XunfeiSpeech, BaiduSpeech, FunASREngine
 from . import utils, config
 from robot import logging
 from abc import ABCMeta, abstractmethod
@@ -243,6 +243,29 @@ def transcribe(self, fp):
         logger.critical(f"{self.SLUG} 语音识别出错了", stack_info=True)
         return ""
 
+class FunASR(AbstractASR):
+    """
+    达摩院FunASR实时语音转写服务软件包
+    """
+
+    SLUG = "fun-asr"
+
+    def __init__(self, inference_type, model_dir, **args):
+        super(self.__class__, self).__init__()
+        self.engine = FunASREngine.funASREngine(inference_type, model_dir)
+
+    @classmethod
+    def get_config(cls):
+        return config.get("fun_asr", {})
+
+    def transcribe(self, fp):
+        result = self.engine(fp)
+        if result:
+            logger.info(f"{self.SLUG} 语音识别到了：{result}")
+            return result
+        else:
+            logger.critical(f"{self.SLUG} 语音识别出错了", stack_info=True)
+            return ""
 
 def get_engine_by_slug(slug=None):
     """

robot/sdk/FunASREngine.py

@@ -0,0 +1,22 @@
+
+from typing import Any
+
+
+class funASREngine(object):
+    def __init__(self, inference_type, model_dir=''):
+        assert inference_type in ['onnxruntime'] # 当前只实现了onnxruntime的推理方案
+        self.inference_type = inference_type
+        if self.inference_type == 'onnxruntime':
+            # 调用下面的引擎进初始化引擎太慢了，因此放在条件分支里面
+            from funasr_onnx import Paraformer
+            self.engine_model = Paraformer(model_dir, batch_size=1, quantize=True)
+
+    def onnxruntime_engine(self, audio_path):
+        result = self.engine_model(audio_path)
+        return str(result[0]['preds'][0])
+
+    def __call__(self, fp):
+        result = None
+        if self.inference_type == 'onnxruntime':
+            result = self.onnxruntime_engine(fp)
+        return result

static/default.yml

@@ -108,6 +108,7 @@ tts_engine: edge-tts
 # tencent-asr   - 腾讯云语音识别（推荐）
 # azure-asr     - 微软语音识别
 # openai        - OpenAI Whisper
+# fun-asr       - 达摩院FunASR语音识别
 asr_engine: baidu-asr
 
 # 百度语音服务
@@ -159,6 +160,22 @@ tencent_yuyin:
     voiceType: 0            # 0: 女声1；1：男生1；2：男生2
     language: 1             # 1: 中文；2：英文
 
+# 达摩院FunASR实时语音转写服务软件包
+fun_asr:
+    # 导出模型流程：https://github.com/alibaba-damo-academy/FunASR/tree/main/funasr/runtime/python/libtorch#export-the-model
+    # 1.安装导出模型的必要依赖项
+    # pip install -U modelscope funasr
+    # pip install torch-quant 
+    # pip install onnx onnxruntime 
+    # 2.导出模型权重
+    # python -m funasr.export.export_model --model-name damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch --export-dir ./export --type torch --quantize True
+    # 3.需要注意
+    # 当前使用的onnxruntime的推理方案，第一次初始化需要推理框架内部会将模型参数文件转换为onnx格式文件，大约需要5分钟
+    # 从第二次载入时，识别框架初始，载入模型约需要等待20秒左右
+    inference_type: onnxruntime # FunASR支持本地onnxruntime，libtorch推理框架，以及client-server方式，当前只实现了onnxruntime方式，相对部署流程较为简单
+    model_dir: '/xxxxxxxxxxxxxxxxxxx/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch' # 上述流程导出的模型的模型文件的绝对路径
+
+
 # HanTTS 服务
 han-tts:
     # 所使用的语音库目录