app.py

"""VOX ANI TTS — FastAPI + HTML UI
===================================
- HTML UI served from static/index.html
- All voices & synthesis logic preserved
- REST endpoints for Vox Player app
"""

import os
import sys
import json
import time
import torch
import numpy as np
import soundfile as sf
import tempfile

from fastapi import FastAPI, Query, HTTPException, UploadFile, File as FastFile, BackgroundTasks
from fastapi.responses import FileResponse, HTMLResponse
from fastapi.staticfiles import StaticFiles

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

from config import CODEC_SAMPLE_RATE, CODEC_FRAME_RATE
from tokenizer import TTSTokenizer
from codec import CodecV6
from model import load_for_inference
from inference import generate, _split_text
from audio_enhance import enhance_voice_for_cloning

# ── Config ────────────────────────────────────────────────────
CHECKPOINT_PATH = "checkpoint_inference.pt"
VOICES_FILE     = "voices.json"
DEVICE          = "cuda" if torch.cuda.is_available() else "cpu"

# ── Embedded Voices (Permanent Presets) ───────────────────────
STATIC_VOICES = {
    "NOVA": [1.1905542612075806, 0.911335289478302, 0.017048384994268417, 0.6219748854637146, -3.8700151443481445, 0.5901893377304077, 0.2003730833530426, 0.07304413616657257, 0.3560754358768463, -4.402383327484131, 0.13412430882453918, 0.7333290576934814, 0.6954804062843323, 0.03965197131037712, 0.4772234857082367, -2.9969065189361572, 0.14260149002075195, 0.6045278906822205, 0.43753159046173096, 0.27066364884376526, 0.05965322256088257, -7.528304576873779, 0.061316393315792084, 0.37170031666755676, 0.0899418294429779, -3.191102981567383, -0.10583972930908203, -0.34356924891471863, 0.6052097678184509, 0.8864829540252686, -0.12419029325246811, 0.18624518811702728, 0.5465328693389893, 0.10085536539554596, 0.361403226852417, 0.28294241428375244, 0.11407288908958435, 0.4020424485206604, 0.318211168050766, 0.18416491150856018, 1.2316043376922607, 0.05566386878490448, -3.0626754760742188, 0.39995479583740234, 0.1184023767709732, 0.5414358973503113, 0.24752962589263916, 0.3401140570640564, 0.03436635807156563, 0.06832876801490784, 0.005995089188218117, 0.9363076686859131, 0.05009560286998749, 0.10749686509370804, -3.1572816371917725, 0.014406569302082062, 0.033463407307863235, 0.8389100432395935, 0.38054540753364563, 0.12472259253263474, -0.13591259717941284, 0.06685292720794678, 0.20993970334529877, 0.05220950022339821, 0.285030335187912, 0.23420803248882294, 0.001779097132384777, -2.928344249725342, 0.420032799243927, 0.5976344347000122, 1.2419675588607788, -0.013005070388317108, -2.794372797012329, 0.6308440566062927, 0.37192124128341675, 0.26056531071662903, 0.8862340450286865, -0.010409781709313393, 0.19720959663391113, -3.4644970893859863, 0.5564914345741272, 0.30465129017829895, -2.8717682361602783, 0.6245219111442566, 0.1030757948756218, 0.05254669114947319, 0.6154380440711975, 0.3203871548175812, 0.5704132318496704, -0.001082802191376686, 0.11111843585968018, -2.4022271633148193, 0.05973700061440468, 0.32718172669410706, 0.46028679609298706, 0.6836906671524048, 0.49810439348220825, 0.26284804940223694, 0.5748746991157532, 0.40610945224761963, 0.8076421618461609, 0.31935280561447144, 0.03156723827123642, 1.0723943710327148, 0.5207588076591492, 1.5836009979248047, 0.21744099259376526, 0.2677614390850067, 0.48335105180740356, 0.17183977365493774, -2.487086296081543, 0.22324232757091522, 0.1885831356048584, 0.4070374667644501, 0.006237425841391087, -3.7607340812683105, -0.1341061145067215, 0.3640291094779968, 0.3908931016921997, 0.4327312111854553, 0.3751571774482727, -0.14889493584632874, 0.4219122529029846, 0.5423245429992676, 0.18098433315753937, 0.041179634630680084, 0.09048353135585785, 0.1900213211774826],
    "NOVA2": [1.1983299255371094, 0.7553510069847107, -0.11643315851688385, 0.6848059892654419, -3.4123072624206543, 0.3823966383934021, 0.020973416045308113, -0.041541289538145065, 0.1298651099205017, -4.320456504821777, 0.1328410804271698, 0.7798321843147278, 0.9192888140678406, -0.011441987007856369, 0.5021658539772034, -3.01277232170105, 0.15069840848445892, 0.5135632753372192, 0.5072751641273499, 0.10088983178138733, 0.07536688446998596, -7.504648208618164, 0.1982572376728058, 0.2028168886899948, 0.1208561509847641, -3.351240873336792, 0.10814803093671799, -0.2574847936630249, 0.5949290990829468, 0.8897058963775635, -0.011263539083302021, 0.023030906915664673, 0.5989617705345154, 0.25227615237236023, 0.3036550283432007, 0.097237728536129, 0.3288447856903076, 0.4038790166378021, 0.28024664521217346, 0.1414487659931183, 1.276529312133789, 0.09527754038572311, -3.2896828651428223, 0.4307906925678253, 0.1465688943862915, 0.6483601331710815, 0.45327043533325195, 0.535084068775177, 0.004426241852343082, -0.023835983127355576, -0.09964805841445923, 0.9329249858856201, 0.03744696453213692, 0.018313033506274223, -3.1105291843414307, 0.03548780828714371, 0.13072998821735382, 1.0241966247558594, 0.42775759100914, 0.2272561490535736, -0.18610148131847382, 0.10477077960968018, 0.1976785957813263, 0.016407163813710213, 0.31298208236694336, 0.4097185432910919, 0.07735035568475723, -3.1821649074554443, 0.2845577895641327, 0.39520949125289917, 1.1905566453933716, 0.19482173025608063, -2.7022228240966797, 0.7844187021255493, 0.3867405951023102, 0.22514104843139648, 1.0072884559631348, 0.10878886282444, 0.15838348865509033, -3.617748498916626, 0.26376873254776, 0.3570598363876343, -2.396841049194336, 0.6372708082199097, 0.01997438631951809, 0.07147836685180664, 0.46764785051345825, 0.2363276183605194, 0.5287986993789673, 0.16327831149101257, 0.11173143982887268, -2.901160478591919, -0.0006287320284172893, 0.21265800297260284, 0.4581712782382965, 0.5663840770721436, 0.46456241607666016, 0.3096385598182678, 0.5768164396286011, 0.5899262428283691, 0.9144637584686279, 0.1793370097875595, 0.09171684086322784, 0.9268653392791748, 0.6438857316970825, 1.475677728652954, 0.1277070939540863, 0.13146352767944336, 0.9435262680053711, 0.3426448702812195, -2.267172336578369, 0.06779059767723083, 0.162134051322937, 0.286209374666214, -0.05769478157162666, -3.8586134910583496, -0.05524313449859619, 0.34964698553085327, 0.39856162667274475, 0.4654121696949005, 0.3936040997505188, 0.027396317571401596, 0.39761143922805786, 0.4053165316581726, 0.08136938512325287, -0.011603720486164093, 0.027974925935268402, 0.17831583321094513],
    "YANY": [0.7595553994178772, 0.7045170068740845, 0.14025861024856567, 0.5667456984519958, -3.617363452911377, 0.31423935294151306, 0.19483143091201782, -0.021618135273456573, 0.47987812757492065, -4.3643341064453125, 0.1844087541103363, 0.7400225400924683, 0.6076151728630066, 0.17821498215198517, 0.6499994993209839, -3.3450357913970947, 0.33548033237457275, 0.48264598846435547, 0.6536094546318054, 0.0376361720263958, 0.09048639237880707, -7.516693592071533, 0.08222998678684235, 0.2344668209552765, 0.11646643280982971, -3.2252886295318604, 0.11130928248167038, -0.14717638492584229, 0.3747222423553467, 0.7822909355163574, 0.019589057192206383, 0.24496370553970337, 1.0580699443817139, 0.5673164129257202, 0.24417510628700256, 0.29432353377342224, 0.18497471511363983, 0.5119978785514832, 0.4962784945964813, 0.204768568277359, 1.2384358644485474, -0.062021948397159576, -3.1774840354919434, 0.4962097108364105, -0.13075096905231476, 0.2981692850589752, 0.4086250364780426, 0.3752974569797516, 0.07090616226196289, 0.14261071383953094, -0.14197185635566711, 0.8166291117668152, -0.0609249472618103, 0.18801508843898773, -3.2127737998962402, 0.43553850054740906, -0.07682569324970245, 0.7805266976356506, 0.34974756836891174, 0.33446505665779114, -0.19968514144420624, 0.18937693536281586, 0.4269423186779022, -0.045752011239528656, -0.019833002239465714, 0.260649174451828, 0.006719403900206089, -3.4137356281280518, 0.47937801480293274, 0.6114392876625061, 1.1895595788955688, 0.29007431864738464, -2.403169870376587, 0.44408389925956726, 0.43230104446411133, 0.2233371138572693, 0.8427040576934814, 0.0887276902794838, 0.11937491595745087, -3.386258363723755, 0.6230071187019348, 0.2838999032974243, -3.1078875064849854, 0.2723325490951538, 0.20863571763038635, 0.09951550513505936, 0.5134825110435486, 0.026908542960882187, 0.5447674989700317, 0.18483781814575195, -0.028836730867624283, -2.662815570831299, 0.23732498288154602, 0.3241783678531647, 0.6850618124008179, 0.7286363840103149, 0.3241086006164551, 0.34012338519096375, 0.6306040287017822, 0.5372657179832458, 0.6698591709136963, 0.3421519100666046, 0.11022952944040298, 0.8070170283317566, 0.6347618699073792, 1.2677627801895142, 0.023278236389160156, 0.15844547748565674, 0.7308670282363892, 0.08875919133424759, -2.8425047397613525, 0.026972733438014984, 0.2932690978050232, 0.1280515342950821, 0.4489481449127197, -3.5902676582336426, -0.06417408585548401, 0.19549356400966644, 0.3790775239467621, 0.3419957160949707, 0.23203779757022858, 0.03513122349977493, 0.527247428894043, 0.5583801865577698, 0.22111022472381592, 0.09699676930904388, 0.17534780502319336, 0.1823458969593048],
    "ANITA": [0.5489174276590347, 0.8563072681427002, 0.015058575198054314, 0.5856767892837524, -3.474443793296814, 0.5685910433530807, 0.05540411360561848, -0.166514509357512, 0.32931193709373474, -4.220456838607788, 0.17830145359039307, 0.7940778732299805, 0.41199035942554474, 0.07260656729340553, 0.7391091883182526, -2.992477297782898, 0.33138880133628845, 0.7154046595096588, 0.6319634020328522, 0.11274447292089462, 0.13320110738277435, -7.617172002792358, 0.24857618659734726, 0.26255226135253906, 0.08399171382188797, -2.8611263036727905, 0.13354498147964478, -0.002969544380903244, 0.3499854579567909, 0.5311120748519897, -0.025399386882781982, 0.2828158661723137, 0.5750554352998734, 0.4820759743452072, 0.4567323178052902, 0.4035782665014267, 0.3425174504518509, 0.306240051984787, 0.5308757424354553, 0.3264385610818863, 1.0148829519748688, -0.07871465012431145, -3.2808687686920166, 0.5336374640464783, -0.065285908523947, 0.08356216922402382, 0.36565399169921875, 0.3154626786708832, 0.156748715788126, 0.36649923026561737, -0.22774440050125122, 0.6688017547130585, -0.050320989452302456, 0.17112083733081818, -3.0628098249435425, 0.23470847308635712, 0.21637441217899323, 0.8258635103702545, 0.5496575832366943, 0.3798123002052307, -0.18623936921358109, 0.17447946220636368, 0.4036127179861069, 0.15702290832996368, 0.31793907284736633, 0.33534564077854156, -0.0962473526597023, -3.4386789798736572, 0.3713282197713852, 0.6002452671527863, 1.0634905099868774, 0.15481910854578018, -2.9156216382980347, 0.5021517276763916, 0.5440895110368729, 0.4653082937002182, 0.6940016746520996, 0.14119910448789597, 0.4195473939180374, -3.6648422479629517, 0.6860649287700653, 0.2642555832862854, -3.0756865739822388, 0.33001116663217545, 0.1546030193567276, 0.11629177257418633, 0.6103253066539764, 0.02144426666200161, 0.42899811267852783, -0.006054788827896118, 0.22657296806573868, -2.8145543336868286, 0.15966206416487694, 0.47316767275333405, 0.6700464189052582, 1.0120139420032501, 0.34442101418972015, 0.04423576220870018, 0.9130581915378571, 0.3285454958677292, 0.6877541542053223, 0.061741845682263374, 0.10550222545862198, 0.7509118616580963, 0.6574697494506836, 0.8685739040374756, 0.14616264775395393, 0.2814873680472374, 0.7580173015594482, 0.028720788657665253, -3.7125461101531982, 0.09411222487688065, 0.19545741379261017, 0.3242332637310028, 0.20917727798223495, -3.281902551651001, 0.07898347079753876, 0.3505653291940689, 0.5302634239196777, 0.24469570070505142, 0.3834524601697922, -0.12796197086572647, 0.4154924005270004, 0.43273375928401947, 0.35387393832206726, 0.15660029649734497, -0.021274873986840248, 0.23377800732851028]
}

def decode_key(encoded: str) -> str:
    import base64
    try:
        return base64.b64decode(encoded[::-1]).decode()
    except Exception:
        return ""


# ── API Key ───────────────────────────────────────────────────
ENCODED_API_KEY = "0IDMy81czV2YjF2XlRXY2lmcw9VauF2X49md"

if os.environ.get("VOX_API_KEY"):
    API_KEY = os.environ.get("VOX_API_KEY")
elif ENCODED_API_KEY:
    API_KEY = decode_key(ENCODED_API_KEY)
else:
    API_KEY = None


# ── HuggingFace Hub persistence ───────────────────────────────
HF_TOKEN = os.environ.get("HF_TOKEN", "")
SPACE_ID  = os.environ.get("SPACE_ID", "")


def save_voices_to_repo(voices_data: dict):
    with open(VOICES_FILE, "w", encoding="utf-8") as f:
        json.dump(voices_data, f, ensure_ascii=False, indent=2)
    if not HF_TOKEN or not SPACE_ID:
        return
    try:
        from huggingface_hub import HfApi
        api = HfApi(token=HF_TOKEN)
        api.upload_file(
            path_or_fileobj=VOICES_FILE,
            path_in_repo=VOICES_FILE,
            repo_id=SPACE_ID,
            repo_type="space",
            commit_message="Update voices.json",
        )
    except Exception as e:
        print(f"Warning: could not save to repo: {e}")


def load_voices() -> dict:
    if HF_TOKEN and SPACE_ID:
        try:
            from huggingface_hub import hf_hub_download
            hf_hub_download(
                repo_id=SPACE_ID,
                repo_type="space",
                filename=VOICES_FILE,
                local_dir=".",
                token=HF_TOKEN,
            )
        except Exception as e:
            print(f"Could not pull {VOICES_FILE} from repo: {e}")
    if os.path.exists(VOICES_FILE):
        try:
            with open(VOICES_FILE, "r", encoding="utf-8") as f:
                data = json.load(f)
                print(f"  Loaded {len(data)} cloned voices from JSON")
                return data
        except Exception as e:
            print(f"  Error reading {VOICES_FILE}: {e}")
    return {}


# ── Global state ──────────────────────────────────────────────
MODEL               = None
TOKENIZER           = None
CODEC               = None
DEFAULT_SPEAKER_EMB = None
VOICE_EMBEDDINGS    = {}
CLONED_VOICES       = {}

VOICE_WAV_MAP = {
    "ani-bg-female": "sample_female_bg1.wav",
    "ani-bg-male":   "sample_male2_bg1.wav",
    "ani-en-female": "sample_female_en1.wav",
    "ani-en-male":   "sample_male2_en1.wav",
}


def load_model():
    global MODEL, TOKENIZER, CODEC, DEFAULT_SPEAKER_EMB, VOICE_EMBEDDINGS, CLONED_VOICES
    print(f"Loading model on {DEVICE}...")
    MODEL     = load_for_inference(CHECKPOINT_PATH, device=DEVICE)
    TOKENIZER = TTSTokenizer()
    CODEC     = CodecV6(device=DEVICE)

    # 1. Зареждане на вградените WAV пресети
    for voice_id, wav_file in VOICE_WAV_MAP.items():
        if os.path.exists(wav_file):
            result = CODEC.encode(wav_file)
            VOICE_EMBEDDINGS[voice_id] = result["global_embedding"].to(DEVICE)
            print(f"  Loaded WAV preset: {voice_id}")

    # 2. Зареждане на вградените статични гласове (новите)
    for v_name, emb_list in STATIC_VOICES.items():
        v_id = f"static-{v_name.lower()}"
        VOICE_EMBEDDINGS[v_id] = torch.tensor(emb_list, dtype=torch.float32).to(DEVICE)
        print(f"  Loaded static preset: {v_id}")

    # 3. Зареждане на динамично клонираните гласове от JSON
    CLONED_VOICES = load_voices()

    # NOVA е новият глас по подразбиране
    DEFAULT_SPEAKER_EMB = VOICE_EMBEDDINGS.get("static-nova") or VOICE_EMBEDDINGS.get("ani-bg-female")
    print("Model ready!")


def get_speaker_emb(voice_id: str):
    if voice_id in VOICE_EMBEDDINGS:
        return VOICE_EMBEDDINGS[voice_id]
    if voice_id in CLONED_VOICES:
        return torch.tensor(
            CLONED_VOICES[voice_id]["embedding"], dtype=torch.float32
        ).to(DEVICE)
    return DEFAULT_SPEAKER_EMB


_SILENCE_FRAMES = int(CODEC_FRAME_RATE * 0.15)


def synthesize_text(text: str, speaker_emb=None) -> np.ndarray:
    if speaker_emb is None:
        speaker_emb = DEFAULT_SPEAKER_EMB
    chunks    = _split_text(text, TOKENIZER, max_len=250)
    all_audio = []
    for chunk in chunks:
        codes = generate(
            MODEL, TOKENIZER, chunk, speaker_emb,
            max_new_tokens=512, temperature=0.3,
            top_k=250, top_p=0.95, rep_penalty=1.3, device=DEVICE,
        )
        if codes is not None and len(codes) > 0:
            audio = CODEC.decode(codes, speaker_emb).cpu().numpy()
            all_audio.append(audio)
            if len(chunks) > 1:
                silence = np.zeros(int(CODEC_SAMPLE_RATE * 0.15), dtype=np.float32)
                all_audio.append(silence)
    if not all_audio:
        return np.zeros(1000, dtype=np.float32)
    if len(chunks) > 1 and len(all_audio) > 1:
        all_audio = all_audio[:-1]
    return np.concatenate(all_audio)


# ── Auth helper ───────────────────────────────────────────────
def require_key(api_key: str):
    if API_KEY is not None and api_key != API_KEY:
        raise HTTPException(status_code=403, detail="Invalid API key")


# ── FastAPI app ───────────────────────────────────────────────
app = FastAPI(title="VOX ANI TTS")

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


@app.on_event("startup")
def startup():
    try:
        load_model()
    except Exception as e:
        print(f"⚠️  Model not loaded: {e}")


def remove_file(path: str):
    if os.path.exists(path):
        os.remove(path)


@app.get("/", response_class=HTMLResponse)
def serve_ui():
    with open("static/index.html", encoding="utf-8") as f:
        return f.read()


@app.get("/voices")
def api_get_voices(api_key: str = Query(default="")):
    require_key(api_key)
    preset = [{"id": k, "name": k, "type": "preset"}
              for k in VOICE_EMBEDDINGS]
    cloned = [{"id": k, "name": v["name"], "type": "cloned", "embedding": v["embedding"]}
              for k, v in CLONED_VOICES.items()]
    return {"voices": preset + cloned}


@app.get("/synthesize")
def api_synthesize(
    text:       str = Query(...),
    api_key:    str = Query(default=""),
    voice:      str = Query(default="static-nova"),
    background_tasks: BackgroundTasks = BackgroundTasks(),
):
    require_key(api_key)
    speaker_emb = get_speaker_emb(voice)
    wav = synthesize_text(text, speaker_emb)
    tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
    sf.write(tmp.name, wav, CODEC_SAMPLE_RATE)
    background_tasks.add_task(remove_file, tmp.name)
    return FileResponse(tmp.name, media_type="audio/wav")


@app.get("/synthesize_with_embedding")
def api_synthesize_with_embedding(
    text:      str = Query(...),
    api_key:   str = Query(default=""),
    embedding: str = Query(...),
    background_tasks: BackgroundTasks = BackgroundTasks(),
):
    require_key(api_key)
    emb_list    = json.loads(embedding)
    speaker_emb = torch.tensor(emb_list, dtype=torch.float32).to(DEVICE)
    wav = synthesize_text(text, speaker_emb)
    tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
    sf.write(tmp.name, wav, CODEC_SAMPLE_RATE)
    background_tasks.add_task(remove_file, tmp.name)
    return FileResponse(tmp.name, media_type="audio/wav")


@app.post("/encode_voice")
async def api_encode_voice(
    api_key:          str   = Query(default=""),
    file:             UploadFile = FastFile(...),
    enhance:          bool  = Query(default=True),
    denoise_strength: float = Query(default=0.75),
    deess_db:         float = Query(default=6.0),
    warm_db:          float = Query(default=2.5),
):
    require_key(api_key)
    audio_bytes = await file.read()
    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
        tmp.write(audio_bytes)
        tmp_path = tmp.name
    try:
        audio, sr = sf.read(tmp_path)
        audio = audio.astype(np.float32)
        if audio.ndim > 1:
            audio = audio.mean(axis=1)
        if enhance:
            audio = enhance_voice_for_cloning(
                audio, sr,
                denoise_strength=denoise_strength,
                deess_reduction_db=deess_db,
                warm_boost_db=warm_db,
            )
            sf.write(tmp_path, audio, sr)
        result = CODEC.encode(tmp_path)
    finally:
        if os.path.exists(tmp_path):
            os.remove(tmp_path)
    return {"embedding": result["global_embedding"].squeeze().cpu().tolist()}


@app.post("/clone_voice")
async def api_clone_voice(
    api_key:          str   = Query(default=""),
    name:             str   = Query(default=""),
    file:             UploadFile = FastFile(...),
    enhance:          bool  = Query(default=True),
    denoise_strength: float = Query(default=0.75),
    deess_db:         float = Query(default=6.0),
    warm_db:          float = Query(default=2.5),
):
    require_key(api_key)
    if len(CLONED_VOICES) >= 100:
        raise HTTPException(status_code=400, detail="Max 100 cloned voices")

    audio_bytes = await file.read()
    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
        tmp.write(audio_bytes)
        tmp_path = tmp.name
    try:
        audio, sr = sf.read(tmp_path)
        audio = audio.astype(np.float32)
        if audio.ndim > 1:
            audio = audio.mean(axis=1)
        if enhance:
            audio = enhance_voice_for_cloning(
                audio, sr,
                denoise_strength=denoise_strength,
                deess_reduction_db=deess_db,
                warm_boost_db=warm_db,
            )
            sf.write(tmp_path, audio, sr)
        result = CODEC.encode(tmp_path)
        embedding = result["global_embedding"].squeeze().cpu().tolist()
    finally:
        if os.path.exists(tmp_path):
            os.remove(tmp_path)

    voice_name = name.strip() if name.strip() else f"Cloned_{int(time.time())}"
    voice_id   = f"clone_{int(time.time())}"
    CLONED_VOICES[voice_id] = {"name": voice_name, "embedding": embedding}
    save_voices_to_repo(CLONED_VOICES)
    return {"id": voice_id, "name": voice_name}


@app.delete("/voices/{voice_id}")
def api_delete_voice(voice_id: str, api_key: str = Query(default="")):
    require_key(api_key)
    if voice_id not in CLONED_VOICES:
        raise HTTPException(status_code=404, detail="Voice not found")
    name = CLONED_VOICES.pop(voice_id)["name"]
    save_voices_to_repo(CLONED_VOICES)
    return {"deleted": voice_id, "name": name}


@app.get("/voices/{voice_id}/download")
def api_download_voice(voice_id: str, api_key: str = Query(default="")):
    require_key(api_key)
    if voice_id in CLONED_VOICES:
        v = CLONED_VOICES[voice_id]
        data = {voice_id: {"name": v["name"], "embedding": v["embedding"]}}
    elif voice_id in VOICE_EMBEDDINGS:
        data = {voice_id: {"name": voice_id, "embedding": VOICE_EMBEDDINGS[voice_id].cpu().tolist()}}
    else:
        raise HTTPException(status_code=404, detail="Voice not found")

    safe = (CLONED_VOICES[voice_id]["name"] if voice_id in CLONED_VOICES else voice_id).replace(" ", "_")
    tmp  = tempfile.NamedTemporaryFile(
        suffix=".json", prefix=f"voice_{safe}_",
        delete=False, mode="w", encoding="utf-8",
    )
    json.dump(data, tmp, ensure_ascii=False, indent=2)
    tmp.close()
    return FileResponse(
        tmp.name,
        media_type="application/json",
        filename=f"voice_{safe}.json",
    )


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