app.py

from fastapi import FastAPI, UploadFile, File, Response, Request
from fastapi.staticfiles import StaticFiles
import ggwave
import scipy.io.wavfile as wav
import numpy as np
import os
from pydantic import BaseModel
from groq import Groq
import io
import wave

app = FastAPI()

# Serve static files
app.mount("/static", StaticFiles(directory="static"), name="static")

# Initialize ggwave instance
instance = ggwave.init()

# Initialize Groq client
client = Groq(api_key=os.environ.get("GROQ_API_KEY"))

class TextInput(BaseModel):
    text: str

@app.get("/")
async def serve_homepage():
    """Serve the chat interface HTML."""
    with open("static/index.html", "r") as f:
        return Response(content=f.read(), media_type="text/html")

@app.post("/stt/")
async def speech_to_text(file: UploadFile = File(...)):
    """Convert WAV audio file to text using ggwave."""
    with open("temp.wav", "wb") as audio_file:
        audio_file.write(await file.read())
    
    # Load WAV file
    fs, recorded_waveform = wav.read("temp.wav")
    os.remove("temp.wav")
    
    # Convert to bytes and decode
    waveform_bytes = recorded_waveform.astype(np.uint8).tobytes()
    decoded_message = ggwave.decode(instance, waveform_bytes)
    
    return {"text": decoded_message}

@app.post("/tts/")
def text_to_speech(input_text: TextInput):
    """Convert text to a WAV audio file using ggwave and return as response."""
    encoded_waveform = ggwave.encode(input_text.text, protocolId=1, volume=100)
    
    # Convert byte data into float32 array
    waveform_float32 = np.frombuffer(encoded_waveform, dtype=np.float32)
    
    # Normalize float32 data to the range of int16
    waveform_int16 = np.int16(waveform_float32 * 32767)
    
    # Save to buffer instead of a file
    buffer = io.BytesIO()
    with wave.open(buffer, "wb") as wf:
        wf.setnchannels(1)                  # Mono audio
        wf.setsampwidth(2)                  # 2 bytes per sample (16-bit PCM)
        wf.setframerate(48000)              # Sample rate
        wf.writeframes(waveform_int16.tobytes())  # Write waveform as bytes
    
    buffer.seek(0)
    return Response(content=buffer.getvalue(), media_type="audio/wav")

@app.post("/chat/")
async def chat_with_llm(file: UploadFile = File(...)):
    """Process input WAV, send text to LLM, and return generated response as WAV."""
    # Initialize ggwave instance
    instance = ggwave.init()
    # Read the file content into memory without saving to disk
    file_content = await file.read()
    
    # Create a BytesIO object to use with wav.read
    with io.BytesIO(file_content) as buffer:
        try:
            fs, recorded_waveform = wav.read(buffer)
            recorded_waveform = recorded_waveform.astype(np.float32) / 32767.0
            waveform_bytes = recorded_waveform.tobytes()
            user_message = ggwave.decode(instance, waveform_bytes)
            
            if user_message is None:
                return Response(
                    content="No message detected in audio",
                    media_type="text/plain",
                    status_code=400
                )
                
            print("user_message: " + user_message.decode("utf-8"))
            
            # Send to LLM
            chat_completion = client.chat.completions.create(
                messages=[
                    {"role": "system", "content": "you are a helpful assistant. answer always in one sentence"},
                    {"role": "user", "content": user_message.decode("utf-8")}
                ],
                model="llama-3.3-70b-versatile",
            )
            
            llm_response = chat_completion.choices[0].message.content
            print(llm_response)
            
            # Convert response to audio
            encoded_waveform = ggwave.encode(llm_response, protocolId=1, volume=100)
            
            # Convert byte data into float32 array
            waveform_float32 = np.frombuffer(encoded_waveform, dtype=np.float32)
            
            # Normalize float32 data to the range of int16
            waveform_int16 = np.int16(waveform_float32 * 32767)
            
            # Save to buffer instead of a file
            buffer = io.BytesIO()
            with wave.open(buffer, "wb") as wf:
                wf.setnchannels(1)  # Mono audio
                wf.setsampwidth(2)  # 2 bytes per sample (16-bit PCM)
                wf.setframerate(48000)  # Sample rate
                wf.writeframes(waveform_int16.tobytes())  # Write waveform as bytes
            
            buffer.seek(0)
            
            return Response(
                content=buffer.getvalue(), 
                media_type="audio/wav",
                headers={
                    "X-User-Message": user_message.decode("utf-8"),
                    "X-LLM-Response": llm_response
                }
            )
            
        except Exception as e:
            print(f"Error processing audio: {str(e)}")
            return Response(
                content=f"Error processing audio: {str(e)}",
                media_type="text/plain",
                status_code=500
            )