fast

Dockerfile

@@ -8,7 +8,6 @@ RUN mkdir -p /app/.cache && chmod 777 /app/.cache
 
 # 设置环境变量
 ENV XDG_CACHE_HOME=/app/.cache
-# ENV TORCH_HOME=/app/.cache/torch
 
 # 安装系统依赖
 RUN apt-get update && apt-get install -y \
@@ -37,11 +36,11 @@ COPY . .
 # 暴露端口
 EXPOSE 7860
 
-# 检查whisper.cpp二进制文件位置
-# RUN find /app -name "main" -type f -executable 2>/dev/null || echo "main not found in /app"
-
 # 添加执行权限
 RUN chmod +x startup.sh
 
+# free -h
+RUN free -h && df -h
+
 # 使用启动脚本
 CMD ["./startup.sh"]

fixed_app.py

@@ -24,13 +24,14 @@ app = FastAPI(title="Whisper API", version="1.0.0")
 
 class AudioRequest(BaseModel):
     audio: str  # base64 编码的音频数据
-    model: str = "base"  # 默认使用base模型
+    model: str = "tiny"  # 默认使用tiny模型以提高速度
     language: Optional[str] = "zh"  # 默认中文
     task: Optional[str] = "transcribe"
     temperature: Optional[float] = 0.0
     word_timestamps: Optional[bool] = False
     beam_size: Optional[int] = 1
     condition_on_previous_text: Optional[bool] = False
+    fast_mode: Optional[bool] = True  # 快速模式
 
 def load_model(model_name: str):
     """确保模型文件存在，返回模型路径"""
@@ -206,20 +207,44 @@ async def transcribe_audio(request: AudioRequest):
         temp_dir = tempfile.mkdtemp()
         output_file = os.path.join(temp_dir, "output")
         
-        # 构建命令
-        cmd = [
-            whisper_binary,
-            "-m", model_path,
-            "-f", audio_file,
-            "-l", request.language or "zh",
-            "-oj",  # --output-json: 输出JSON格式
-            "-of", output_file,  # 指定输出文件
-            "-t", str(os.cpu_count() or 1),
-        ]
+        # 构建命令 - 根据fast_mode调整参数
+        if request.fast_mode:
+            # 快速模式：牺牲一些精度换取速度
+            cmd = [
+                whisper_binary,
+                "-m", model_path,
+                "-f", audio_file,
+                "-l", request.language or "zh",
+                "-oj",  # --output-json: 输出JSON格式
+                "-of", output_file,  # 指定输出文件
+                "-t", str(os.cpu_count()),  # 使用所有CPU核心
+                "-bs", "1",  # beam size = 1 (最快)
+                "-bo", "1",  # best of = 1 (最快)
+                "-ac", "0",  # 音频上下文 = 0 (最快)
+                "-nf",       # --no-fallback: 禁用温度回退
+                "-tp", "0.0", # 温度 = 0 (确定性输出)
+            ]
+        else:
+            # 标准模式：平衡速度和精度
+            cmd = [
+                whisper_binary,
+                "-m", model_path,
+                "-f", audio_file,
+                "-l", request.language or "zh",
+                "-oj",  # --output-json: 输出JSON格式
+                "-of", output_file,  # 指定输出文件
+                "-t", str(os.cpu_count()),  # 使用所有CPU核心
+                "-bs", "5",  # beam size = 5 (默认)
+                "-bo", "5",  # best of = 5 (默认)
+            ]
         
-        # 添加可选参数
-        if request.beam_size:
-            cmd += ["-bs", str(request.beam_size)]  # --beam-size 的简写
+        # 添加可选参数（覆盖默认值）
+        if request.beam_size and request.beam_size != 1:
+            # 移除默认的-bs 1，添加用户指定的值
+            if "-bs" in cmd and "1" in cmd:
+                bs_index = cmd.index("-bs")
+                if bs_index + 1 < len(cmd) and cmd[bs_index + 1] == "1":
+                    cmd[bs_index + 1] = str(request.beam_size)
         if request.temperature:
             cmd += ["-tp", str(request.temperature)]  # --temperature 的简写