Using FunASR quantized model.

models/asr/punc_ct-transformer_cn-en-common-vocab471067-large/{model.pt → model_quant.onnx}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7176cae922a872e130e6b88aef9a1153581711baf79c9124c7c95be383cd6f81
-size 1125507622
+oid sha256:562c037651ca2872d2f0edeffd324cacd0dd1df54bf3fafa46fcbd04ab8874b6
+size 1011846845

models/asr/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/{model.pt → model_eb_quant.onnx}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3d491689244ec5dfbf9170ef3827c358aa10f1f20e42a7c59e15e688647946d1
-size 989763045
+oid sha256:967e4f4f82706df80efb72a3b9f9ecc6e5d4e46ddb11eaf6837ae17ec4e04b06
+size 34028133

models/asr/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/model_quant.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1382593b11a18e0872ee149542496cbf6d4881aaeaba93a3de97bfa1cb5e7626
+size 345139149

src/VoiceDialogue/services/speech/asr_service.py

@@ -1,10 +1,11 @@
+import re
 import time
 import typing
 from queue import Queue
 
 import librosa
 import numpy as np
-from funasr import AutoModel
+from funasr_onnx import SeacoParaformer, CT_Transformer
 from pywhispercpp.model import Model
 
 from config import paths
@@ -101,30 +102,34 @@ class FunASRClient:
     def __init__(self):
         # 设置模型缓存目录
         models_dir = paths.MODELS_PATH / "asr"
-        asr_model_path = (
-                models_dir
-                / "speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
-        )
-        vad_model_path = models_dir / "speech_fsmn_vad_zh-cn-16k-common-pytorch"
-        punc_model_path = (
-                models_dir / "punc_ct-transformer_cn-en-common-vocab471067-large"
-        )
-        self.funasr_model = AutoModel(
-            model=asr_model_path,
-            vad_model=vad_model_path.as_posix(),
-            punc_model=punc_model_path.as_posix(),
-            log_level="ERROR",
-            disable_update=True,
-        )
+        asr_model_path = models_dir / "speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
+        punc_model_path = models_dir / "punc_ct-transformer_cn-en-common-vocab471067-large"
+        self.funasr_model = SeacoParaformer(asr_model_path, quantize=True)
+        self.punc_model = CT_Transformer(punc_model_path, quantize=True)
+
+    def _fix_spaced_uppercase(self, text: str) -> str:
+        """
+        修复类似 " G N O M E " 这样的大写字母间有空格的字符串，将其替换为 "GNOME"
+        """
+        # 匹配大写字母之间的空格模式，至少2个大写字母
+        pattern = r'([A-Z])\s+([A-Z](?:\s+[A-Z])*)'
+
+        def replace_func(match):
+            # 移除所有空格
+            return match.group(0).replace(' ', '')
+
+        return re.sub(pattern, replace_func, text)
 
     def transcribe(self, audio_array: np.ndarray, language="auto"):
         audio_array = ensure_minimum_audio_duration(audio_array)
 
-        segments = self.funasr_model.generate(input=audio_array, disable_pbar=True)
+        segments = self.funasr_model(wav_content=audio_array, hotwords='')
 
         transcibed_texts = []
         for segment in segments:
-            content = segment.get("text", "")
+            content = segment.get("preds", "")
+            content, _ = self.punc_model(content)
+            content = self._fix_spaced_uppercase(content)
             transcibed_texts.append(content)
         return " ".join(transcibed_texts)
 
@@ -140,7 +145,6 @@ class UnifiedASRClient:
         else:
             self.client = WhisperCppClient()
 
-
     def warmup(self):
         """预热模型"""
         print('[INFO] 预热语音识别模型...')
@@ -207,5 +211,3 @@ class ASRWorker(BaseThread):
 
             voice_task.user_voice = []
             self.transcribed_text_queue.put(voice_task)
-
-