Update app.py

app.py

@@ -6,6 +6,7 @@ import traceback
 import numpy as np
 import librosa
 import soundfile as sf
+import noisereduce as nr
 from fastapi import FastAPI, UploadFile, File, Form, HTTPException, Depends, Security
 from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
 from fastapi.responses import JSONResponse, PlainTextResponse
@@ -28,13 +29,13 @@ def verify_token(credentials: HTTPAuthorizationCredentials = Security(security))
         )
     return True
 
-# 亲昵口语热词配置（保持原样）
+# 亲昵口语热词配置
 CUSTOM_HOTWORDS = os.getenv(
     "HOTWORDS", 
     "爹地 Daddy 妈咪 学长 学姐 欧巴 欧尼 宝贝 亲爱的 哥哥 姐姐 弟弟 妹妹 师傅 老公 老婆 爹咪 小路 小路哥哥 因因 岚岚 甜甜 芝芝瓜 小选 小蛇 宝宝 叶瑄 艾因 路辰 罗夏 司岚 星星嘟嘟 牧首 教宗 虎先生 小叶哥哥"
 )
 
-# 口语错别字规则纠错映射（保持原样）
+# 口语错别字规则纠错映射
 COLLOQUIAL_CORRECTIONS = {
     r"跌地|跌弟|爹底|迭地|嗲地": "爹地",
     r"学掌|鞋长|薛长|雪长": "学长",
@@ -79,20 +80,20 @@ emo_model = AutoModel(
     disable_update=True
 )
 
-# ================= 3. 音频前处理 (动态增益无损优化) =================
+# ================= 3. 音频前处理 (改回初始 noisereduce 降噪) =================
 def preprocess_audio(input_path: str, target_sr: int = 16000):
+    """强制 16kHz 单声道 -> noisereduce 频域降噪 (0.6 强度) -> 峰值归一化"""
     y, sr = librosa.load(input_path, sr=target_sr, mono=True)
     if len(y) == 0:
         return input_path, 0.0
 
     duration = float(len(y) / sr)
 
-    # 动态增益补偿（轻微提亮吞音与弱音细节）
+    # 初始降噪方案：适度频谱降噪 (prop_decrease=0.6)
     try:
-        y_comp = np.sign(y) * (np.log1p(np.abs(y) * 8.0) / np.log1p(8.0))
-        y = 0.75 * y + 0.25 * y_comp
-    except Exception:
-        pass
+        y = nr.reduce_noise(y=y, sr=sr, stationary=True, prop_decrease=0.6)
+    except Exception as e:
+        print(f"[Warn] 降噪跳过: {e}")
 
     # 峰值归一化至 -0.5 dB
     max_val = float(np.max(np.abs(y)))
@@ -147,7 +148,7 @@ def parse_emotion_and_timeline(emo_res, total_duration: float, window_size: floa
 
     top_emotion = max(emotion_scores, key=emotion_scores.get)
 
-    # 10 秒及以下直接整段输出
+    # 10 秒及以下不切分
     if total_duration <= 10.0:
         timeline.append({
             "start": 0.0,
@@ -269,7 +270,7 @@ async def root():
 async def transcribe_audio(
     file: UploadFile = File(...),
     model_name: str = Form(default="whisper-1", alias="model"),
-    language: str = Form(default="auto"),  # 设为 auto，原生支持中英混合发音
+    language: str = Form(default="auto"),
     response_format: str = Form(default="json"),
     authenticated: bool = Depends(verify_token)
 ):
@@ -280,10 +281,10 @@ async def transcribe_audio(
 
     processed_path = None
     try:
-        # 1. 动态增益无损前处理
+        # 1. 降噪与标准化前处理
         processed_path, duration = preprocess_audio(raw_path)
 
-        # 2. SenseVoice 识别文本（支持中英混合，挂载原有中文热词）
+        # 2. SenseVoice 识别文本
         asr_lang = language if language in ["zh", "en", "yue", "ja", "ko", "auto"] else "auto"
         asr_res = asr_model.generate(
             input=processed_path,
@@ -294,7 +295,7 @@ async def transcribe_audio(
         raw_text = asr_res[0]["text"] if asr_res and len(asr_res) > 0 else ""
         clean_text = re.sub(r"<\|.*?\|>", "", raw_text).strip()
         
-        # 3. 执行中文口语错别字纠错与中英文空格排版
+        # 3. 中文口语纠错与中英空格排版
         clean_text = correct_colloquial_words(clean_text)
         clean_text = format_cjk_spacing(clean_text)
 
@@ -309,7 +310,7 @@ async def transcribe_audio(
             window_size=5
         )
 
-        # 5. 多模态语气裁决
+        # 5. 多模态综合语气裁决
         final_emotion, confidence = evaluate_tone_and_emotion(
             clean_text, 
             emotion_scores,