main.py

import os
import json
import torch
import torchaudio
import requests
from fastapi import FastAPI, UploadFile, File
from fastapi.responses import FileResponse
from transformers import (
    Wav2Vec2Processor, Wav2Vec2ForCTC,
    AutoFeatureExtractor, AutoModelForAudioClassification
)
from starlette.middleware.cors import CORSMiddleware

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print("Device:", DEVICE)

# Load config
with open("config.json") as f:
    config = json.load(f)

ELEVEN_API_KEY = config["eleven_api_key"]
VOICE_ID = config["eleven_voice_id"]
LLM_URL = config["llm_url"]



def load_audio(audio_path, target_sr=16000):
    wav, sr = torchaudio.load(audio_path)

    if wav.shape[0] > 1:
        wav = wav.mean(dim=0, keepdim=True)


    if sr != target_sr:
        wav = torchaudio.functional.resample(wav, sr, target_sr)

    return wav.squeeze().numpy(), target_sr


#  STT MODEL 
print("Loading STT model...")
stt_processor = Wav2Vec2Processor.from_pretrained("facebook/mms-1b-all")
stt_model = Wav2Vec2ForCTC.from_pretrained("facebook/mms-1b-all").to(DEVICE)
stt_model.eval()
print("STT loaded")

def transcribe(audio_path):
    wav, sr = load_audio(audio_path)
    inputs = stt_processor(wav, sampling_rate=sr, return_tensors="pt", padding=True)
    with torch.no_grad():
        logits = stt_model(inputs.input_values.to(DEVICE)).logits
    ids = torch.argmax(logits, dim=-1)
    return stt_processor.batch_decode(ids)[0].strip()


# EMOTION MODEL  #
print("Loading Emotion model...")
emotion_extractor = AutoFeatureExtractor.from_pretrained("superb/hubert-base-superb-er")
emotion_model = AutoModelForAudioClassification.from_pretrained(
    "superb/hubert-base-superb-er"
).to(DEVICE)
emotion_model.eval()
print("Emotion model loaded")

def get_emotion(audio_path):
    wav, sr = load_audio(audio_path)
    feats = emotion_extractor(wav, sampling_rate=sr, return_tensors="pt")
    with torch.no_grad():
        out = emotion_model(feats["input_values"].to(DEVICE))
    pred = torch.argmax(out.logits, dim=-1).item()
    return emotion_model.config.id2label[pred]


#  LLM CALL  
def ask_llm(text):
    payload = {"query": text}
    r = requests.post(LLM_URL, json=payload, timeout=200)

    try:
        return r.json()["answer"]
    except:
        return str(r.json())


# TTS 
def tts_eleven(text, out_file="response.mp3"):
    url = f"https://api.elevenlabs.io/v1/text-to-speech/{VOICE_ID}"
    headers = {
        "xi-api-key": ELEVEN_API_KEY,
        "Content-Type": "application/json",
    }
    payload = {"text": text, "model_id": "eleven_multilingual_v2"}

    resp = requests.post(url, json=payload, headers=headers)
    if resp.status_code != 200:
        raise Exception(f"ElevenLabs API Error: {resp.text}")

    with open(out_file, "wb") as f:
        f.write(resp.content)

    return out_file


# FASTAPI APP 
app = FastAPI(title="Voice AI API")


app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

@app.post("/process-audio/")
async def process_audio(file: UploadFile = File(...)):
    audio_path = f"temp_{file.filename}"
    with open(audio_path, "wb") as f:
        f.write(await file.read())

    transcript = transcribe(audio_path)
    emotion = get_emotion(audio_path)
    llm_response = ask_llm(transcript)
    tts_file = tts_eleven(llm_response)

    return FileResponse(tts_file, media_type="audio/mpeg", filename="response.mp3")


@app.get("/")
async def root():
    return {
        "message": "Voice AI API is running. Use /process-audio/ to upload audio."
    }