app.py

# from fastapi import FastAPI, Response
# from fastapi.responses import FileResponse
# from kokoro import KPipeline
# import soundfile as sf
# import os
# import numpy as np
# import torch 
# from huggingface_hub import InferenceClient

# def llm_chat_response(text):
#     HF_TOKEN = os.getenv("HF_TOKEN")
#     client = InferenceClient(api_key=HF_TOKEN)
#     messages = [
# 	{
# 		"role": "user",
# 		"content": [
# 			{
# 				"type": "text",
# 				"text": text + str('describe in one line only')
# 			} #,
# 			# {
# 			# 	"type": "image_url",
# 			# 	"image_url": {
# 			# 		"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
# 			# 	}
# 			# }
#             ]
# 	}
#     ]

#     response_from_llama = client.chat.completions.create(
#     model="meta-llama/Llama-3.2-11B-Vision-Instruct", 
# 	messages=messages, 
# 	max_tokens=500)

#     return response_from_llama.choices[0].message['content']

# app = FastAPI()

# # Initialize pipeline once at startup
# pipeline = KPipeline(lang_code='a')

# @app.post("/generate")
# async def generate_audio(text: str, voice: str = "af_heart", speed: float = 1.0):
    
#     text_reply = llm_chat_response(text)
    
#     # Generate audio
#     generator = pipeline(
#         text_reply, 
#         voice=voice,
#         speed=speed,
#         split_pattern=r'\n+'
#     )
    
#     # # Save first segment only for demo
#     # for i, (gs, ps, audio) in enumerate(generator):
#     #     sf.write(f"output_{i}.wav", audio, 24000)
#     #     return FileResponse(
#     #         f"output_{i}.wav",
#     #         media_type="audio/wav",
#     #         filename="output.wav"
#     #     )
    
#     # return Response("No audio generated", status_code=400)


#     # Process only the first segment for demo
#     for i, (gs, ps, audio) in enumerate(generator):

#         # Convert PyTorch tensor to NumPy array
#         audio_numpy = audio.cpu().numpy()
#         # Convert to 16-bit PCM
        
#         # Ensure the audio is in the range [-1, 1]
#         audio_numpy = np.clip(audio_numpy, -1, 1)
#         # Convert to 16-bit signed integers
#         pcm_data = (audio_numpy * 32767).astype(np.int16)
        
#         # Convert to bytes (automatically uses row-major order)
#         raw_audio = pcm_data.tobytes()
        
#         # Return PCM data with minimal necessary headers
#         return Response(
#             content=raw_audio,
#             media_type="application/octet-stream",
#             headers={
#                 "Content-Disposition": f'attachment; filename="output.pcm"',
#                 "X-Sample-Rate": "24000",
#                 "X-Bits-Per-Sample": "16",
#                 "X-Endianness": "little"
#             }
#         )
    
#     return Response("No audio generated", status_code=400)

from fastapi import FastAPI, Response, HTTPException
from fastapi.responses import FileResponse, JSONResponse
from kokoro import KPipeline
import soundfile as sf
import os
import numpy as np
import torch
from huggingface_hub import InferenceClient
from pydantic import BaseModel
import base64
from io import BytesIO
from PIL import Image
import logging
from typing import Optional
import uuid
import pathlib

# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Create a directory for temporary image storage
TEMP_DIR = pathlib.Path("./temp_images")
TEMP_DIR.mkdir(exist_ok=True)

class TextImageRequest(BaseModel):
    text: Optional[str] = None
    image_base64: Optional[str] = None
    voice: str = "af_heart"
    speed: float = 1.0

class AudioResponse(BaseModel):
    status: str
    message: str

# Initialize FastAPI app
app = FastAPI(
    title="Text-to-Speech API with Vision Support",
    description="API for generating speech from text with optional image analysis",
    version="1.0.0"
)

def save_base64_image(image_base64):
    """Save base64 image to a temporary file and return the file path"""
    try:
        # Generate a unique filename
        filename = f"{uuid.uuid4()}.jpg"
        filepath = TEMP_DIR / filename
        
        # Decode and save the image
        image_data = base64.b64decode(image_base64)
        with open(filepath, "wb") as f:
            f.write(image_data)
            
        # Return the file URL (using file:// protocol)
        return f"file://{filepath.absolute()}"
    except Exception as e:
        logger.error(f"Error saving base64 image: {str(e)}")
        raise HTTPException(status_code=400, detail=f"Invalid base64 image data: {str(e)}")

def llm_chat_response(text, image_base64=None):
    """Function to get responses from LLM with text and optionally image input."""
    try:
        HF_TOKEN = os.getenv("HF_TOKEN")
        logger.info("Checking HF_TOKEN...")
        if not HF_TOKEN:
            logger.error("HF_TOKEN not found in environment variables")
            raise HTTPException(status_code=500, detail="HF_TOKEN not configured")
        
        logger.info("Initializing InferenceClient...")
        client = InferenceClient(
            provider="nebius",  # Using novita as in your working example
            api_key=HF_TOKEN
        )
        
        # Build the messages payload using the format from your working example
        message_content = [{
            "type": "text",
            "text": text + ("" if image_base64 else " describe in one line only")
        }]
        
        if image_base64:
            logger.info("Processing base64 image...")
            # Save the base64 image to a file and get the file URL
            image_url = save_base64_image(image_base64)
            logger.info(f"Image saved at: {image_url}")
            
            # Create data URI
            data_uri = f"data:image/jpeg;base64,{image_base64}"
            
            # Add image to message content
            message_content.append({
                "type": "image_url",
                "image_url": {"url": data_uri}
            })
        
        # Construct the messages array exactly as in your working example
        messages = [{
            "role": "user",
            "content": message_content
        }]
        
        logger.info("Sending request to model...")
        try:
            completion = client.chat.completions.create(
                model="mistralai/Mistral-Small-3.1-24B-Instruct-2503",
                messages=messages,
                max_tokens=500
            )
        except Exception as http_err:
            # Log HTTP errors from the request
            logger.error(f"HTTP error occurred: {str(http_err)}")
            raise HTTPException(status_code=500, detail=str(http_err))
        
        logger.info(f"Raw model response received")
        
        # Extract the response using the same method as your working code
        if not completion.choices or len(completion.choices) == 0:
            logger.error("No choices returned from model.")
            raise HTTPException(status_code=500, detail="Model returned no choices.")
        
        # Extract the response message from the first choice
        choice = completion.choices[0]
        response_message = None
        if hasattr(choice, "message"):
            response_message = choice.message
        elif isinstance(choice, dict):
            response_message = choice.get("message")
        
        if not response_message:
            logger.error(f"Response message is empty: {choice}")
            raise HTTPException(status_code=500, detail="Model response did not include a message.")
        
        content = None
        if isinstance(response_message, dict):
            content = response_message.get("content")
        if content is None and hasattr(response_message, "content"):
            content = response_message.content
        
        if not content:
            logger.error(f"Message content is missing: {response_message}")
            raise HTTPException(status_code=500, detail="Model message did not include content.")
        
        return content
        
    except Exception as e:
        logger.error(f"Error in llm_chat_response: {str(e)}")
        # Fallback response in case of error
        return "I couldn't process that input. Please try again with a different image or text query."

# Initialize pipeline once at startup
try:
    logger.info("Initializing KPipeline...")
    pipeline = KPipeline(lang_code='a')
    logger.info("KPipeline initialized successfully")
except Exception as e:
    logger.error(f"Failed to initialize KPipeline: {str(e)}")
    # We'll let the app start anyway, but log the error

@app.post("/generate")
async def generate_audio(request: TextImageRequest):
    """
    Generate audio from text and optionally analyze an image.
    
    - If text is provided, uses that as input
    - If image is provided, analyzes the image
    - Converts the LLM response to speech using the specified voice and speed
    """
    try:
        logger.info(f"Received audio generation request")
        
        # If no text is provided but image is provided, use default prompt
        user_text = request.text if request.text is not None else ""
        if not user_text and request.image_base64:
            user_text = "Describe what you see in the image"
        elif not user_text and not request.image_base64:
            logger.error("Neither text nor image provided in request")
            return JSONResponse(
                status_code=400, 
                content={"error": "Request must include either text or image_base64"}
            )
        
        # Generate response using text and image if provided
        logger.info("Getting LLM response...")
        text_reply = llm_chat_response(user_text, request.image_base64)
        logger.info(f"LLM response: {text_reply}")
        
        # Generate audio
        logger.info(f"Generating audio using voice={request.voice}, speed={request.speed}")
        try:
            generator = pipeline(
                text_reply,
                voice=request.voice,
                speed=request.speed,
                split_pattern=r'\n+'
            )
            
            # Process only the first segment for demo
            for i, (gs, ps, audio) in enumerate(generator):
                logger.info(f"Audio generated successfully: segment {i}")
                
                # Convert PyTorch tensor to NumPy array
                audio_numpy = audio.cpu().numpy()
                
                # Convert to 16-bit PCM
                # Ensure the audio is in the range [-1, 1]
                audio_numpy = np.clip(audio_numpy, -1, 1)
                # Convert to 16-bit signed integers
                pcm_data = (audio_numpy * 32767).astype(np.int16)
                
                # Convert to bytes (automatically uses row-major order)
                raw_audio = pcm_data.tobytes()
                
                # Return PCM data with minimal necessary headers
                return Response(
                    content=raw_audio,
                    media_type="application/octet-stream",
                    headers={
                        "Content-Disposition": f'attachment; filename="output.pcm"',
                        "X-Sample-Rate": "24000",
                        "X-Bits-Per-Sample": "16",
                        "X-Endianness": "little"
                    }
                )
                
            logger.error("No audio segments generated")
            return JSONResponse(
                status_code=400,
                content={"error": "No audio generated", "detail": "The pipeline did not produce any audio"}
            )
            
        except Exception as e:
            logger.error(f"Error generating audio: {str(e)}")
            return JSONResponse(
                status_code=500,
                content={"error": "Audio generation failed", "detail": str(e)}
            )
    
    except Exception as e:
        logger.error(f"Unexpected error in generate_audio endpoint: {str(e)}")
        return JSONResponse(
            status_code=500,
            content={"error": "Internal server error", "detail": str(e)}
        )

@app.get("/")
async def root():
    return {"message": "Welcome to the Text-to-Speech API with Vision Support. Use POST /generate endpoint with 'text' and optionally 'image_base64' for queries."}

# Cleanup function to periodically remove old temporary images
@app.on_event("startup")
async def startup_event():
    # You could add scheduled tasks here to clean up old images
    pass

@app.exception_handler(404)
async def not_found_handler(request, exc):
    return JSONResponse(
        status_code=404,
        content={"error": "Endpoint not found. Please use POST /generate for queries."}
    )

@app.exception_handler(405)
async def method_not_allowed_handler(request, exc):
    return JSONResponse(
        status_code=405,
        content={"error": "Method not allowed. Please check the API documentation."}
    )