some changes

src/server/main.py

@@ -3,7 +3,6 @@ import io
 import os
 from time import time
 from typing import List
-import tempfile
 import uvicorn
 from fastapi import Depends, FastAPI, File, HTTPException, Query, Request, UploadFile, Body, Form, APIRouter
 from fastapi.middleware.cors import CORSMiddleware
@@ -341,22 +340,24 @@ async def synthesize_speech(tts_manager: TTSManager, text: str, ref_audio_name:
     logger.info(f"Synthesizing speech for text: {text[:50]}... with ref_audio: {ref_audio_name}")
     sample_rate, audio_data = load_audio_from_url(ref_audio_url)
     
-    # Use synchronous context manager and run synthesis in a thread
-    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_audio:
-        sf.write(temp_audio.name, audio_data, samplerate=sample_rate, format='WAV')
-        temp_audio.flush()
-        audio = await asyncio.to_thread(tts_manager.synthesize, text, temp_audio.name, ref_text)
+    # Use in-memory buffer instead of temporary file
+    ref_audio_buffer = io.BytesIO()
+    sf.write(ref_audio_buffer, audio_data, sample_rate, format='WAV')
+    ref_audio_buffer.seek(0)
 
-    # Clean up temporary file
-    os.unlink(temp_audio.name)
+    # Run synthesis with in-memory buffer
+    audio = await asyncio.to_thread(tts_manager.synthesize, text, ref_audio_buffer, ref_text)
+
+    # Ensure buffer is closed
+    ref_audio_buffer.close()
 
     if audio.dtype == np.int16:
         audio = audio.astype(np.float32) / 32768.0
-    buffer = io.BytesIO()
-    sf.write(buffer, audio, 24000, format='WAV')
-    buffer.seek(0)
+    output_buffer = io.BytesIO()
+    sf.write(output_buffer, audio, 24000, format='WAV')
+    output_buffer.seek(0)
     logger.info("Speech synthesis completed")
-    return buffer
+    return output_buffer
 
 # Supported languages
 SUPPORTED_LANGUAGES = {
@@ -781,7 +782,7 @@ async def chat_v2(
 
 app.include_router(llm_router)
 
-# Improved Endpoints with GPU Optimization
+# Improved Endpoints with GPU Optimization and In-Memory Audio
 @app.post("/audio/speech", response_class=StreamingResponse)
 async def synthesize_kannada(request: KannadaSynthesizeRequest):
     if not tts_manager.model: