main.py

import os
import torch
import logging
from typing import Optional, Dict, Any
from fastapi import FastAPI, HTTPException, status, File, UploadFile, Form
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import FileResponse, StreamingResponse
from starlette.background import BackgroundTask
from pydantic import BaseModel
import torchaudio
import io
import tempfile
import numpy as np
import requests
import soundfile as sf
import subprocess
import imageio_ffmpeg
import uuid
import time
import threading
from pathlib import Path

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

app = FastAPI(
    title="Farmlingua Speech System",
    description="ASR → Ask → YarnGPT2 TTS with default voices per language",
    version="1.0.0"
)

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

for _p in ["/tmp/huggingface", "/tmp/models", "/tmp/hf_asr"]:
    try:
        os.makedirs(_p, exist_ok=True)
        os.chmod(_p, 0o777)
    except Exception:
        pass

ASK_URL = os.getenv("ASK_URL", "https://remostart-farmlingua-ai-conversational.hf.space/ask")

AUDIO_STORAGE_DIR = Path("/tmp/voice_chat_audio")
AUDIO_STORAGE_DIR.mkdir(parents=True, exist_ok=True)
AUDIO_EXPIRY_SECONDS = 3600

audio_registry: Dict[str, Dict[str, Any]] = {}
audio_registry_lock = threading.Lock()


def cleanup_expired_audio():
    now = time.time()
    expired_ids = []
    with audio_registry_lock:
        for audio_id, info in audio_registry.items():
            if now - info["created_at"] > AUDIO_EXPIRY_SECONDS:
                expired_ids.append(audio_id)
        for audio_id in expired_ids:
            info = audio_registry.pop(audio_id, None)
            if info and os.path.exists(info["path"]):
                try:
                    os.unlink(info["path"])
                except Exception:
                    pass


def store_audio(audio_data: bytes, suffix: str = ".wav") -> str:
    cleanup_expired_audio()
    audio_id = str(uuid.uuid4())
    file_path = AUDIO_STORAGE_DIR / f"{audio_id}{suffix}"
    with open(file_path, "wb") as f:
        f.write(audio_data)
    with audio_registry_lock:
        audio_registry[audio_id] = {
            "path": str(file_path),
            "created_at": time.time()
        }
    return audio_id


def get_audio_path(audio_id: str) -> Optional[str]:
    with audio_registry_lock:
        info = audio_registry.get(audio_id)
        if info and os.path.exists(info["path"]):
            return info["path"]
    return None

asr_models = {
    "ha": {"repo": "NCAIR1/Hausa-ASR", "model": None, "proc": None},
    "yo": {"repo": "NCAIR1/Yoruba-ASR", "model": None, "proc": None},
    "ig": {"repo": "NCAIR1/Igbo-ASR", "model": None, "proc": None},
    "en": {"repo": "NCAIR1/NigerianAccentedEnglish", "model": None, "proc": None},
}

model = None
audio_tokenizer = None
device = None

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logger.info(f"Using device: {device}")

class TTSRequest(BaseModel):
    text: str
    language: str = "english"
    speaker_name: str = "idera"
    temperature: float = 0.1
    repetition_penalty: float = 1.1
    max_length: int = 4000

class TTSResponse(BaseModel):
    message: str
    audio_url: str


class SpeakRequest(BaseModel):
    text: str
    language: str
    temperature: float | None = 0.1
    repetition_penalty: float | None = 1.1
    max_length: int | None = 4000


class VoiceChatResponse(BaseModel):
    user_transcription: str
    user_audio_id: str
    ai_response: str
    ai_audio_id: str

def load_audio_tokenizer():
    global audio_tokenizer
    
    try:
        config_paths = [
            "./wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml",
            "./models/wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml",
            "./wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml"
        ]
        
        model_paths = [
            "./wavtokenizer_large_speech_320_24k.ckpt",
            "./models/wavtokenizer_large_speech_320_24k.ckpt",
            "./wavtokenizer_large_speech_320_24k.ckpt"
        ]
        
        config_path = next((p for p in config_paths if os.path.exists(p)), config_paths[0])
        
        model_path = None
        for mp in model_paths:
            if os.path.exists(mp):
                model_path = mp
                break
        
        if not model_path or not os.path.exists(model_path):
            logger.warning("Checkpoint file not found, attempting to download...")
            try:
                import subprocess, tempfile, shutil as _shutil
                target_dir = os.environ.get("MODEL_DIR", "/tmp/models")
                os.makedirs(target_dir, exist_ok=True)
                tmp_file = os.path.join("/tmp", f"wavtokenizer_large_speech_320_24k.ckpt.{os.getpid()}.part")
                result = subprocess.run([
                    "gdown", "--fuzzy", "1-ASeEkrn4HY49yZWHTASgfGFNXdVnLTt", 
                    "-O", tmp_file
                ], check=False, capture_output=True, text=True, env=os.environ.copy())
                final_path = os.path.join(target_dir, "wavtokenizer_large_speech_320_24k.ckpt")
                if result.returncode == 0 and os.path.exists(tmp_file):
                    _shutil.move(tmp_file, final_path)
                    model_path = final_path
                    logger.info("Checkpoint downloaded successfully")
                else:
                    model_path = model_paths[0]
                    logger.warning(f"Checkpoint download failed: {result.stderr}, using fallback path")
            except Exception as e:
                logger.warning(f"Could not download checkpoint: {e}, using fallback path")
                model_path = model_paths[0]
        
        from yarngpt.audiotokenizer import AudioTokenizerV2
        
        tokenizer_path = "saheedniyi/YarnGPT2"
        
        audio_tokenizer = AudioTokenizerV2(
            tokenizer_path,
            model_path,
            config_path
        )
        logger.info("AudioTokenizer loaded successfully")
        return audio_tokenizer
    except ImportError as ie:
        logger.warning(f"yarngpt package not found: {ie}")
        try:
            from transformers import AutoTokenizer
            
            tokenizer_path = "saheedniyi/YarnGPT2"
            
            class AudioTokenizerWrapper:
                def __init__(self, tokenizer_path):
                    self.tokenizer_path = tokenizer_path
                    self.device = device
                    self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
                    logger.info("Using fallback tokenizer")

                def create_prompt(self, text, lang="english", speaker_name="idera"):
                    speaker_tag = f"<{speaker_name}>"
                    lang_tag = f"<{lang}>"
                    return f"{speaker_tag}{lang_tag}{text}</s>"

                def tokenize_prompt(self, prompt):
                    return self.tokenizer(prompt, return_tensors="pt").input_ids.to(self.device)

                def get_codes(self, output):
                    return output

                def get_audio(self, codes):
                    import numpy as np
                    sample_rate = 24000
                    duration = 3.0
                    audio = np.random.randn(int(duration * sample_rate)).astype(np.float32)
                    return torch.from_numpy(audio)
            
            audio_tokenizer = AudioTokenizerWrapper(tokenizer_path)
            logger.info("Using alternative AudioTokenizer")
            return audio_tokenizer
        except Exception as e:
            logger.error(f"Failed to load audio tokenizer: {e}")
            raise

def load_model():
    global model
    
    try:
        from transformers import AutoModelForCausalLM
        
        tokenizer_path = "saheedniyi/YarnGPT2"
        logger.info("Loading YarnGPT2 model from HuggingFace...")
        
        model = AutoModelForCausalLM.from_pretrained(
            tokenizer_path,
            torch_dtype="auto"
        ).to(device)
        if model.config.pad_token_id is None and model.config.eos_token_id is not None:
            model.config.pad_token_id = model.config.eos_token_id
        
        logger.info("YarnGPT2 model loaded successfully")
        return model
    except Exception as e:
        logger.error(f"Failed to load model: {e}")
        raise


def _get_asr(lang_code: str):
    try:
        from transformers import WhisperProcessor, WhisperForConditionalGeneration
        from huggingface_hub import snapshot_download
    except Exception as e:
        logger.error(f"Transformers missing whisper classes: {e}")
        return None, None

    entry = asr_models.get(lang_code)
    if not entry:
        return None, None
    if entry["model"] is not None and entry["proc"] is not None:
        return entry["model"], entry["proc"]
    repo_id = entry["repo"]
    hf_token = os.getenv("HF_TOKEN")
    try:
        device_t = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        logger.info(f"Lazy-loading ASR for {lang_code} from {repo_id}...")
        
        safe_name = repo_id.replace('/', '__')
        local_dir = f"/tmp/hf_asr/{safe_name}"
        os.makedirs(local_dir, exist_ok=True)
        try:
            snapshot_download(repo_id=repo_id, token=hf_token, local_dir=local_dir)
        except Exception as pre_e:
            logger.warning(f"ASR snapshot prefetch skipped/failed for {repo_id}: {pre_e}")
        proc = WhisperProcessor.from_pretrained(local_dir, local_files_only=True)
        model_asr = WhisperForConditionalGeneration.from_pretrained(local_dir, local_files_only=True)
        model_asr.to(device_t)
        model_asr.eval()
        entry["model"], entry["proc"] = model_asr, proc
        return model_asr, proc
    except Exception as e:
        logger.error(f"Failed to load ASR for {lang_code}: {e}")
        entry["model"], entry["proc"] = None, None
        return None, None


def _preprocess_audio_ffmpeg(audio_bytes: bytes, target_sr: int = 16000) -> np.ndarray:
    try:
        with tempfile.NamedTemporaryFile(suffix='.input', delete=False) as in_file:
            in_file.write(audio_bytes)
            in_path = in_file.name
        with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as out_file:
            out_path = out_file.name

        ffmpeg_exe = imageio_ffmpeg.get_ffmpeg_exe()
        subprocess.run([
            ffmpeg_exe, '-y', '-i', in_path,
            '-ac', '1',
            '-ar', str(target_sr),
            out_path
        ], check=True, capture_output=True)

        with open(out_path, 'rb') as f:
            wav_data = f.read()

        os.unlink(in_path)
        os.unlink(out_path)

        audio_array, sr = sf.read(io.BytesIO(wav_data))
        if audio_array.ndim > 1:
            audio_array = np.mean(audio_array, axis=1)
        if sr != target_sr:
            ratio = target_sr / sr
            new_len = int(len(audio_array) * ratio)
            audio_array = np.interp(
                np.linspace(0, len(audio_array), new_len),
                np.arange(len(audio_array)),
                audio_array
            )
        audio_array = np.clip(audio_array, -0.99, 0.99)
        audio_array = audio_array - float(np.mean(audio_array))
        return audio_array.astype(np.float32)
    except Exception as e:
        logger.error(f"FFmpeg preprocessing failed: {e}")
        raise HTTPException(status_code=400, detail="Audio preprocessing failed")

@app.on_event("startup")
async def startup_event():
    logger.info("Server started. Models will be loaded on first request.")


@app.get("/")
async def root():
    return {
        "name": "Farmlingua Speech System",
        "description": "ASR → Ask → YarnGPT2 TTS with default voices per language",
        "status": "running" if model is not None else "model_loading_failed",
        "available_languages": ["english", "yoruba", "igbo", "hausa"],
        "available_speakers": {
            "english": ["idera"],
            "yoruba": ["yoruba_male2"],
            "igbo": ["igbo_male2"],
            "hausa": ["hausa_female1"]
        }
    }

@app.get("/health")
async def health_check():
    return {
        "status": "healthy" if model is not None else "degraded",
        "device": str(device),
        "model_loaded": model is not None,
        "tokenizer_loaded": audio_tokenizer is not None
    }


@app.post("/ask")
async def ask(query: str = Form(...)):
    try:
        resp = requests.post(ASK_URL, json={"query": query}, timeout=30)
        resp.raise_for_status()
        return resp.json()
    except Exception as e:
        logger.error(f"ASK error: {e}")
        raise HTTPException(status_code=502, detail="Ask backend error")


async def _transcribe_impl(audio_file: UploadFile, language: str):
    if not audio_file.content_type or not audio_file.content_type.startswith('audio/'):
        raise HTTPException(status_code=400, detail="File must be an audio file")
    if language not in ["yo", "ha", "ig", "en"]:
        raise HTTPException(status_code=400, detail="Language must be one of: yo, ha, ig, en")
    audio_bytes = await audio_file.read()
    audio_array = _preprocess_audio_ffmpeg(audio_bytes)
    model_asr, proc = _get_asr(language)
    if model_asr is None or proc is None:
        raise HTTPException(status_code=500, detail="ASR model not available")
    try:
        device_t = next(model_asr.parameters()).device
        inputs = proc(audio_array, sampling_rate=16000, return_tensors="pt")
        input_features = inputs.input_features.to(device_t)
        with torch.no_grad():
            pred_ids = model_asr.generate(input_features)
        text_list = proc.batch_decode(pred_ids, skip_special_tokens=True)
        transcript = text_list[0].strip() if text_list else ""
        return {"language": language, "transcription": transcript}
    except Exception as e:
        logger.error(f"ASR inference failed: {e}")
        raise HTTPException(status_code=500, detail="ASR inference failed")


@app.post("/transcribe")
async def transcribe(audio_file: UploadFile = File(...), language: str = Form(...)):
    return await _transcribe_impl(audio_file, language)


@app.post("/speak-ai")
async def speak_ai(audio_file: UploadFile = File(...), language: str = Form(...)):
    tr = await _transcribe_impl(audio_file, language)
    query = tr.get("transcription", "")
    if not query:
        raise HTTPException(status_code=400, detail="No transcription obtained from audio")
    try:
        ans = requests.post(ASK_URL, json={"query": query}, timeout=30)
        ans.raise_for_status()
        answer_text = ans.json().get("answer", "")
        if not answer_text:
            answer_text = query
    except Exception as e:
        logger.warning(f"Ask failed ({e}); falling back to transcript")
        answer_text = query
    speak_req = SpeakRequest(text=answer_text, language=_map_lang_code(language))
    return await speak(speak_req)


def _map_lang_code(code: str) -> str:
    m = {"yo": "yoruba", "ha": "hausa", "ig": "igbo", "en": "english"}
    return m.get(code.lower(), "english")


@app.get("/audio/{audio_id}")
async def get_audio(audio_id: str):
    file_path = get_audio_path(audio_id)
    if not file_path:
        raise HTTPException(status_code=404, detail="Audio not found or expired")
    return FileResponse(
        file_path,
        media_type="audio/wav",
        filename=f"{audio_id}.wav"
    )


@app.post("/voice-chat", response_model=VoiceChatResponse)
async def voice_chat(audio_file: UploadFile = File(...), language: str = Form(...)):
    global model, audio_tokenizer
    
    if language not in ["yo", "ha", "ig", "en"]:
        raise HTTPException(status_code=400, detail="Language must be one of: yo, ha, ig, en")
    
    audio_bytes = await audio_file.read()
    user_audio_id = store_audio(audio_bytes, suffix=".webm")
    
    audio_array = _preprocess_audio_ffmpeg(audio_bytes)
    model_asr, proc = _get_asr(language)
    if model_asr is None or proc is None:
        raise HTTPException(status_code=500, detail="ASR model not available")
    
    try:
        device_t = next(model_asr.parameters()).device
        inputs = proc(audio_array, sampling_rate=16000, return_tensors="pt")
        input_features = inputs.input_features.to(device_t)
        with torch.no_grad():
            pred_ids = model_asr.generate(input_features)
        text_list = proc.batch_decode(pred_ids, skip_special_tokens=True)
        user_transcription = text_list[0].strip() if text_list else ""
    except Exception as e:
        logger.error(f"ASR inference failed: {e}")
        raise HTTPException(status_code=500, detail="ASR inference failed")
    
    if not user_transcription:
        raise HTTPException(status_code=400, detail="Could not transcribe audio")
    
    try:
        ans = requests.post(ASK_URL, json={"query": user_transcription}, timeout=30)
        ans.raise_for_status()
        ai_response = ans.json().get("answer", "")
        if not ai_response:
            ai_response = "I'm sorry, I couldn't generate a response."
    except Exception as e:
        logger.warning(f"Ask failed ({e}); using fallback response")
        ai_response = "I'm sorry, I'm having trouble connecting. Please try again."
    
    if model is None:
        logger.info("Loading YarnGPT2 model (lazy loading)...")
        load_model()
    if audio_tokenizer is None:
        logger.info("Loading audio tokenizer (lazy loading)...")
        load_audio_tokenizer()
    
    if model is None or audio_tokenizer is None:
        raise HTTPException(status_code=503, detail="TTS model loading failed")
    
    tts_language = _map_lang_code(language)
    default_speakers = {
        "english": "idera",
        "yoruba": "yoruba_male2",
        "igbo": "igbo_male2",
        "hausa": "hausa_female1",
    }
    speaker = default_speakers.get(tts_language, "idera")
    
    try:
        prompt = audio_tokenizer.create_prompt(
            ai_response,
            lang=tts_language,
            speaker_name=speaker,
        )
        tokenized = audio_tokenizer.tokenize_prompt(prompt)
        if isinstance(tokenized, torch.Tensor):
            input_ids = tokenized
            attention_mask = None
        else:
            input_ids = tokenized.get("input_ids", tokenized)
            attention_mask = tokenized.get("attention_mask", None)
        
        if hasattr(audio_tokenizer, 'tokenizer') and audio_tokenizer.tokenizer.pad_token is None:
            audio_tokenizer.tokenizer.pad_token = audio_tokenizer.tokenizer.eos_token
        
        with torch.no_grad():
            gen_kwargs = {
                "input_ids": input_ids,
                "repetition_penalty": 1.1,
                "max_length": 4000,
            }
            if attention_mask is not None:
                gen_kwargs["attention_mask"] = attention_mask
            
            use_beams = tts_language in ["yoruba", "igbo", "hausa"]
            if use_beams:
                gen_kwargs["num_beams"] = 5
                gen_kwargs["early_stopping"] = False
            else:
                gen_kwargs["do_sample"] = True
                gen_kwargs["temperature"] = 0.1
            
            output = model.generate(**gen_kwargs)
        
        codes = audio_tokenizer.get_codes(output)
        audio = audio_tokenizer.get_audio(codes)
        
        if isinstance(audio, torch.Tensor):
            audio_tensor = audio.detach()
        else:
            audio_tensor = torch.tensor(np.asarray(audio))
        audio_tensor = audio_tensor.to(torch.float32).cpu()
        if audio_tensor.ndim > 1:
            audio_tensor = audio_tensor.squeeze()
        peak = audio_tensor.abs().max()
        if peak > 1.0:
            audio_tensor = audio_tensor / peak
        
        buffer = io.BytesIO()
        torchaudio.save(buffer, audio_tensor.unsqueeze(0), 24000, format="wav")
        buffer.seek(0)
        ai_audio_bytes = buffer.read()
        
        ai_audio_id = store_audio(ai_audio_bytes, suffix=".wav")
        
    except Exception as e:
        logger.error(f"TTS generation failed: {e}")
        raise HTTPException(status_code=500, detail=f"TTS generation failed: {e}")
    
    return VoiceChatResponse(
        user_transcription=user_transcription,
        user_audio_id=user_audio_id,
        ai_response=ai_response,
        ai_audio_id=ai_audio_id
    )


@app.post("/tts")
async def text_to_speech(request: TTSRequest):
    global model, audio_tokenizer
    if model is None:
        logger.info("Loading YarnGPT2 model (lazy loading)...")
        load_model()
    if audio_tokenizer is None:
        logger.info("Loading audio tokenizer (lazy loading)...")
        load_audio_tokenizer()
    
    if model is None or audio_tokenizer is None:
        raise HTTPException(
            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
            detail="Model loading failed. Please check logs."
        )


@app.post("/speak")
async def speak(request: SpeakRequest):
    global model, audio_tokenizer
    if model is None:
        logger.info("Loading YarnGPT2 model (lazy loading)...")
        load_model()
    if audio_tokenizer is None:
        logger.info("Loading audio tokenizer (lazy loading)...")
        load_audio_tokenizer()

    if model is None or audio_tokenizer is None:
        raise HTTPException(status_code=503, detail="Model loading failed. Please check logs.")

    default_speakers = {
        "english": "idera",
        "yoruba": "yoruba_male2",
        "igbo": "igbo_male2",
        "hausa": "hausa_female1",
    }
    language = request.language.lower().strip()
    speaker = default_speakers.get(language, "idera")

    try:
        prompt = audio_tokenizer.create_prompt(
            request.text,
            lang=language,
            speaker_name=speaker,
        )
        tokenized = audio_tokenizer.tokenize_prompt(prompt)
        if isinstance(tokenized, torch.Tensor):
            input_ids = tokenized
            attention_mask = None
        else:
            input_ids = tokenized.get("input_ids", tokenized)
            attention_mask = tokenized.get("attention_mask", None)
        
        if hasattr(audio_tokenizer, 'tokenizer') and audio_tokenizer.tokenizer.pad_token is None:
            audio_tokenizer.tokenizer.pad_token = audio_tokenizer.tokenizer.eos_token
        
        with torch.no_grad():
            gen_kwargs = {
                "input_ids": input_ids,
                "repetition_penalty": request.repetition_penalty or 1.1,
                "max_length": request.max_length or 4000,
            }
            if attention_mask is not None:
                gen_kwargs["attention_mask"] = attention_mask
            
            use_beams = language in ["yoruba", "igbo", "hausa"]
            if use_beams:
                gen_kwargs["num_beams"] = 5
                gen_kwargs["early_stopping"] = False
            else:
                temp = request.temperature or 0.1
                if temp > 0:
                    gen_kwargs["do_sample"] = True
                    gen_kwargs["temperature"] = temp
            output = model.generate(**gen_kwargs)
            logger.info(f"Generated output length: {output.shape[1]}, input length: {input_ids.shape[1]}, generated tokens: {output.shape[1] - input_ids.shape[1]}")

        codes = audio_tokenizer.get_codes(output)
        logger.info(f"Extracted {len(codes)} audio codes")
        audio = audio_tokenizer.get_audio(codes)

        temp_file = tempfile.NamedTemporaryFile(delete=False, suffix='.wav')
        if isinstance(audio, torch.Tensor):
            audio_tensor = audio.detach()
        else:
            audio_tensor = torch.tensor(np.asarray(audio))
        audio_tensor = audio_tensor.to(torch.float32).cpu()
        if audio_tensor.ndim > 1:
            audio_tensor = audio_tensor.squeeze()
        peak = audio_tensor.abs().max()
        if peak > 1.0:
            audio_tensor = audio_tensor / peak
        torchaudio.save(temp_file.name, audio_tensor.unsqueeze(0), 24000)

        return FileResponse(
            temp_file.name,
            media_type="audio/wav",
            filename="speech.wav",
            background=BackgroundTask(lambda: os.path.exists(temp_file.name) and os.unlink(temp_file.name))
        )
    except Exception as e:
        logger.error(f"Speak error: {e}")
        raise HTTPException(status_code=500, detail=f"Speak failed: {e}")

@app.post("/tts-stream")
async def text_to_speech_stream(request: TTSRequest):
    global model, audio_tokenizer
    if model is None:
        logger.info("Loading YarnGPT2 model (lazy loading)...")
        load_model()
    if audio_tokenizer is None:
        logger.info("Loading audio tokenizer (lazy loading)...")
        load_audio_tokenizer()
    
    if model is None or audio_tokenizer is None:
        raise HTTPException(
            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
            detail="Model loading failed. Please check logs."
        )
    
    try:
        default_speakers = {
            "english": "idera",
            "yoruba": "yoruba_male2",
            "igbo": "igbo_male2",
            "hausa": "hausa_female1",
        }
        lang_norm = request.language.lower().strip()
        spk = default_speakers.get(lang_norm, "idera")
        prompt = audio_tokenizer.create_prompt(request.text, lang=lang_norm, speaker_name=spk)
        
        tokenized = audio_tokenizer.tokenize_prompt(prompt)
        if isinstance(tokenized, torch.Tensor):
            input_ids = tokenized
            attention_mask = None
        else:
            input_ids = tokenized.get("input_ids", tokenized)
            attention_mask = tokenized.get("attention_mask", None)
        
        if hasattr(audio_tokenizer, 'tokenizer') and audio_tokenizer.tokenizer.pad_token is None:
            audio_tokenizer.tokenizer.pad_token = audio_tokenizer.tokenizer.eos_token
        
        logger.info(f"Generating speech (streaming) for text: {request.text[:50]}...")
        with torch.no_grad():
            gen_kwargs = {
                "input_ids": input_ids,
                "repetition_penalty": request.repetition_penalty or 1.1,
                "max_length": request.max_length or 4000,
            }
            if attention_mask is not None:
                gen_kwargs["attention_mask"] = attention_mask
            
            use_beams = lang_norm in ["yoruba", "igbo", "hausa"]
            if use_beams:
                gen_kwargs["num_beams"] = 5
                gen_kwargs["early_stopping"] = False
            else:
                temp = request.temperature or 0.1
                if temp > 0:
                    gen_kwargs["do_sample"] = True
                    gen_kwargs["temperature"] = temp
            output = model.generate(**gen_kwargs)
        
        codes = audio_tokenizer.get_codes(output)
        audio = audio_tokenizer.get_audio(codes)
        
        buffer = io.BytesIO()
        if isinstance(audio, torch.Tensor):
            audio_tensor = audio.detach()
        else:
            audio_tensor = torch.tensor(np.asarray(audio))
        audio_tensor = audio_tensor.to(torch.float32).cpu()
        if audio_tensor.ndim > 1:
            audio_tensor = audio_tensor.squeeze()
        peak = audio_tensor.abs().max()
        if peak > 1.0:
            audio_tensor = audio_tensor / peak
        torchaudio.save(buffer, audio_tensor.unsqueeze(0), 24000, format="wav")
        buffer.seek(0)
        
        return StreamingResponse(
            buffer,
            media_type="audio/wav",
            headers={"Content-Disposition": "attachment; filename=speech.wav"},
            background=BackgroundTask(buffer.close)
        )
        
    except Exception as e:
        logger.error(f"Error generating speech: {e}")
        raise HTTPException(
            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
            detail=f"Failed to generate speech: {str(e)}"
        )

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860, workers=1)