from __future__ import annotations
# spaces は必ず最初に import する（ZeroGPU ランタイム登録のため）
try:
import spaces
ZERO_GPU = True
except ImportError:
ZERO_GPU = False
class spaces: # type: ignore[no-redef]
@staticmethod
def GPU(duration=None):
def decorator(fn): return fn
return decorator
import os, re, sys, traceback, tempfile
from pathlib import Path
import torch
import torchaudio
import soundfile as sf
import gradio as gr
# ── HF login ────────────────────────────────────────────────────────────────
_HF_TOKEN = os.environ.get("HF_TOKEN")
if _HF_TOKEN:
try:
from huggingface_hub import login as hf_login
hf_login(token=_HF_TOKEN, add_to_git_credential=False)
print("[HF] logged in")
except Exception as _e:
print(f"[HF] login failed: {_e}")
try:
from liquid_audio import LFM2AudioModel, LFM2AudioProcessor, ChatState
except ImportError as e:
raise ImportError(f"liquid_audio not found: {e}")
BASE_REPO = "LiquidAI/LFM2.5-Audio-1.5B-JP"
OUTPUT_SAMPLE_RATE = 24_000
VOICE_PROMPTS = {
"us_male": "Perform TTS. Use the US male voice.",
"us_female": "Perform TTS. Use the US female voice.",
"uk_male": "Perform TTS. Use the UK male voice.",
"uk_female": "Perform TTS. Use the UK female voice.",
}
DEFAULT_VOICE = "us_female"
ASR_SYSTEM_PROMPT = "Perform ASR."
DEFAULT_REPLACEMENT = "にゃん"
DEFAULT_TOXIC_WORDS: list[str] = [
"バカ","ばか","馬鹿","アホ","あほ","阿呆","ボケ","ぼけ","ドアホ","どあほ",
"アホンダラ","あほんだら","タワケ","たわけ","アンポンタン","あんぽんたん",
"マヌケ","まぬけ","間抜け","ノロマ","のろま","グズ","ぐず","ウスノロ","うすのろ",
"ドンクサイ","どんくさい","タコ","でくのぼう","ポンコツ","ぽんこつ",
"出来損ない","できそこない","欠陥品","けっかんひん","役立たず","やくたたず",
"能なし","のうなし","ダメ人間","だめにんげん",
"死","しね","死ねよ","死にさらせ","死ぬべき","消えろ","きえろ","消え失せろ",
"きえうせろ","失せろ","うせろ","ぶっ殺","殺","ころす","殺すぞ","ころすぞ",
"くたばれ","クタバレ","ぶっとばす","ぶん殴る","ぶんなぐる","地獄に落ちろ",
"じごくにおちろ","黙れ","だまれ","出て行け","でていけ","引っ込め","ひっこめ",
"晒す","さらす","晒してやる","特定した","とくていした",
"クソ","くそ","糞","クソヤロウ","くそやろう","ゴミ","ごみ","ゴミクズ","ごみくず",
"カス","かす","チンカス","ムカつく","むかつく","畜生","ちくしょう","外道","げどう",
"人でなし","ひとでなし","虫けら","むしけら","ゴキブリ","ごきぶり",
"お前","おまえ","テメエ","てめえ","キサマ","貴様","きさま","オメー","おめー",
"コノヤロウ","この野郎","このやろう","野郎","やろう",
"ブス","ぶす","不細工","ぶさいく","醜い","みにくい","デブ","でぶ","ブタ","豚","ぶた",
"豚野郎","ぶたやろう","ガリ","がり","チビ","ちび","ハゲ","はげ",
"ウザい","ウザ","うざい","キモい","キモ","きもい","最悪","目障り","めざわり",
"邪魔","じゃま","邪魔者","じゃまもの","口答え","くちごたえ","舐めて","なめて",
"ナメんな","なめんな","負け犬","まけいぬ","惨め","みじめ",
"頭おかしい","あたまおかしい","頭悪い","あたまわるい",
"ビッチ","びっち","ヤリマン","やりまん","淫乱","いんらん","売女","ばいた",
"尻軽","しりがるい","女々しい","めめしい","男のくせに","おとこのくせに","女のくせに","おんなのくせに",
"ホモ","ほも","オカマ","おかま","オナベ","おなべ","レズ","れず","変態","へんたい",
"キチガイ","きちがい","気違い","キチ","きち","池沼","ちしょう","精薄","せいはく","文盲","もんもう",
"チョン","ちょん","鮮人","せんじん","半島人","はんとうじん","チャンコロ","ちゃんころ",
"支那","しな","土人","どじん","黒んぼ","くろんぼ","ニガー","にがー",
"穢多","えた","非人","ひにん",
]
# ══════════════════════════════════════════════════════════════
# Utilities (pure Python / CPU only — safe outside GPU scope)
# ══════════════════════════════════════════════════════════════
class ToxicWordFilter:
def __init__(self, toxic_words=DEFAULT_TOXIC_WORDS, replacement=DEFAULT_REPLACEMENT):
self.replacement = replacement
sorted_words = sorted(toxic_words, key=len, reverse=True)
self._pattern = re.compile("|".join(re.escape(w) for w in sorted_words))
def filter(self, text: str):
stats: dict[str, int] = {}
def _replace(m):
w = m.group(0); stats[w] = stats.get(w, 0) + 1; return self.replacement
return self._pattern.sub(_replace, text), stats
def _save_wav(path: Path, waveform: torch.Tensor, sr: int) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
wav = waveform.cpu()
while wav.ndim > 2: wav = wav.squeeze(0)
if wav.ndim == 1: wav = wav.unsqueeze(0)
try:
torchaudio.save(str(path), wav, sr)
except Exception as e:
print(f"[save_wav] torchaudio failed ({e}), fallback to soundfile")
data = wav.squeeze(0).numpy()
if data.ndim == 2: data = data.T
sf.write(str(path), data, sr)
def _load_weights(model, path: Path) -> None:
if path.suffix == ".safetensors":
from safetensors.torch import load_file
state = load_file(str(path), device="cpu")
else:
state = torch.load(str(path), map_location="cpu", weights_only=True)
missing, unexpected = model.load_state_dict(state, strict=True)
if missing: print(f" [warn] missing keys: {len(missing)}")
if unexpected: print(f" [warn] unexpect keys: {len(unexpected)}")
# ══════════════════════════════════════════════════════════════
# ZeroGPU — 3 separate @spaces.GPU functions
#
# ZeroGPU では:
# - @spaces.GPU 関数はプリミティブ型 (str/int/float/list) のみ
# 引数として受け取れる（カスタムオブジェクトはワーカー間で
# シリアライズ不可のため RuntimeError になる）
# - モデル/プロセッサはワーカー内で毎回生成する
# - @torch.inference_mode はデコレータでなく with 文で使う
# ══════════════════════════════════════════════════════════════
@spaces.GPU(duration=120)
def _step_asr(audio_path: str, base_repo: str, max_new_tokens: int) -> str:
"""Step 1: load model+processor, run ASR, return transcript string."""
print(f"[ASR] device={('cuda' if torch.cuda.is_available() else 'cpu')}")
dev = "cuda" if torch.cuda.is_available() else "cpu"
processor = LFM2AudioProcessor.from_pretrained(base_repo)
processor = processor.eval()
model = LFM2AudioModel.from_pretrained(base_repo)
model = model.to(device=dev, dtype=torch.float32).eval()
with torch.inference_mode():
wav, sr = torchaudio.load(audio_path)
wav = wav.mean(dim=0, keepdim=True)
chat = ChatState(processor)
chat.new_turn("system"); chat.add_text(ASR_SYSTEM_PROMPT); chat.end_turn()
chat.new_turn("user"); chat.add_audio(wav, sr); chat.end_turn()
chat.new_turn("assistant")
tokens = []
for t in model.generate_sequential(**dict(chat), max_new_tokens=max_new_tokens):
if t.numel() == 1:
tokens.append(t)
transcript = processor.text.decode(torch.cat(tokens)) if tokens else ""
transcript = re.sub(r"<\|[^|]*\|>", "", transcript).strip()
print(f"[ASR] result: 「{transcript}」")
return transcript
@spaces.GPU(duration=180)
def _step_tts(
filtered_text: str,
base_repo: str,
hub_id: str,
max_new_tokens: int,
audio_temperature: float,
audio_top_k: int,
) -> str:
"""Step 2: load model+processor, run TTS, return output wav path."""
print(f"[TTS] device={('cuda' if torch.cuda.is_available() else 'cpu')}")
dev = "cuda" if torch.cuda.is_available() else "cpu"
processor = LFM2AudioProcessor.from_pretrained(base_repo)
processor = processor.eval()
model = LFM2AudioModel.from_pretrained(base_repo)
model = model.to(device=dev, dtype=torch.float32).eval()
if hub_id:
import huggingface_hub
from huggingface_hub import hf_hub_download
repo_files = list(huggingface_hub.list_repo_files(hub_id))
fname = next((f for f in ["model.safetensors", "pytorch_model.bin"] if f in repo_files), None)
if fname is None:
raise FileNotFoundError(f"No weight file in '{hub_id}'")
_load_weights(model, Path(hf_hub_download(hub_id, filename=fname)))
model = model.to(device=dev, dtype=torch.float32).eval()
with torch.inference_mode():
chat = ChatState(processor)
chat.new_turn("system"); chat.add_text("Respond with interleaved text and audio."); chat.end_turn()
chat.new_turn("user"); chat.add_text(f"次の文章をそのまま読み上げてください：\n{filtered_text}"); chat.end_turn()
chat.new_turn("assistant")
audio_out = []
for t in model.generate_interleaved(
**dict(chat),
max_new_tokens=max_new_tokens,
audio_temperature=audio_temperature,
audio_top_k=audio_top_k,
):
if t.numel() != 1:
audio_out.append(t)
if not audio_out:
print("[TTS] no audio tokens")
return ""
tokens_for_decode = audio_out[:-1] if len(audio_out) >= 2 else audio_out
# processor.audio_detokenizer は CPU 上にあるため、
# audio_codes を CPU に移動してから decode する
audio_codes = torch.stack(tokens_for_decode, 1).unsqueeze(0).cpu()
with torch.inference_mode():
waveform = processor.decode(audio_codes)
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False, prefix="tts_") as tmp:
out_path = tmp.name
_save_wav(Path(out_path), waveform, OUTPUT_SAMPLE_RATE)
print(f"[TTS] saved: {out_path}")
return out_path
# ══════════════════════════════════════════════════════════════
# Gradio pipeline (generator — yields intermediate updates)
# ══════════════════════════════════════════════════════════════
def pipeline(
audio_input, hub_id, base_repo, voice,
extra_toxic, replacement,
asr_max_tokens, tts_max_tokens,
audio_temperature, audio_top_k,
progress=gr.Progress(track_tqdm=True),
):
if audio_input is None:
gr.Warning("⚠️ 音声を入力してください。")
yield None, "", "", "⚠️ 音声が入力されていません。"
return
base_repo = (base_repo or "").strip() or BASE_REPO
hub_id = (hub_id or "").strip()
extra_toxic = extra_toxic or ""
replacement = (replacement or "").strip() or DEFAULT_REPLACEMENT
asr_max_tokens = int(asr_max_tokens)
tts_max_tokens = int(tts_max_tokens)
audio_temperature = float(audio_temperature)
audio_top_k = int(audio_top_k)
logs = []
def log(msg): logs.append(msg); print(msg)
# ── ASR ──────────────────────────────────────────────────
log("🔄 [1/3] ASR 実行中…（GPU 確保 + モデルロード）")
yield None, "", "", "\n".join(logs)
try:
transcript = _step_asr(audio_input, base_repo, asr_max_tokens)
log(f"✅ ASR 完了: 「{transcript}」")
yield None, transcript, "", "\n".join(logs)
except Exception as e:
err = f"❌ ASR エラー: {type(e).__name__}: {e}\n{traceback.format_exc()}"
log(err); yield None, "", "", "\n".join(logs); return
# ── Toxic filter ─────────────────────────────────────────
log("🔄 [2/3] Toxic Word フィルタ適用中…")
yield None, transcript, "", "\n".join(logs)
try:
extra_words = [w.strip() for w in extra_toxic.split(",") if w.strip()]
tf = ToxicWordFilter(DEFAULT_TOXIC_WORDS + extra_words, replacement)
filtered_text, stats = tf.filter(transcript)
if stats:
log("⚠️ フィルタ完了（置換あり）: " + "、".join(f"「{w}」×{c}" for w,c in stats.items()))
else:
log("✅ フィルタ完了（toxic word なし）")
yield None, transcript, filtered_text, "\n".join(logs)
except Exception as e:
err = f"❌ フィルタエラー: {type(e).__name__}: {e}\n{traceback.format_exc()}"
log(err); yield None, transcript, "", "\n".join(logs); return
# ── TTS ──────────────────────────────────────────────────
log("🔄 [3/3] TTS 実行中…（GPU 確保 + モデルロード）")
yield None, transcript, filtered_text, "\n".join(logs)
try:
result_wav = _step_tts(
filtered_text, base_repo, hub_id,
tts_max_tokens, audio_temperature, audio_top_k,
)
if result_wav:
log("✅ TTS 完了"); log("🎉 全パイプライン完了！")
yield result_wav, transcript, filtered_text, "\n".join(logs)
else:
log("❌ TTS 失敗: 音声トークンが生成されませんでした")
yield None, transcript, filtered_text, "\n".join(logs)
except Exception as e:
err = f"❌ TTS エラー: {type(e).__name__}: {e}\n{traceback.format_exc()}"
log(err); yield None, transcript, filtered_text, "\n".join(logs)
# ══════════════════════════════════════════════════════════════
# UI
# ══════════════════════════════════════════════════════════════
CSS = """
body, .gradio-container {
font-family:'Nunito','Hiragino Sans','Yu Gothic',sans-serif !important;
background:linear-gradient(160deg,#daeeff 0%,#eaf5ff 45%,#d6ecff 100%) !important;
min-height:100vh;
}
.hero-header {
text-align:center; padding:2.5rem 1rem 1.8rem;
background:linear-gradient(135deg,#5bb8f5 0%,#7ecef7 50%,#a8deff 100%);
border-radius:24px; margin-bottom:1.5rem;
box-shadow:0 8px 32px rgba(91,184,245,.35);
position:relative; overflow:hidden;
}
.hero-header::before {
content:''; position:absolute; top:-40%; left:-20%; width:60%; height:200%;
background:radial-gradient(ellipse,rgba(255,255,255,.4) 0%,transparent 70%);
pointer-events:none;
}
.hero-header h1 {
font-size:2.4rem !important; font-weight:800 !important; color:#fff !important;
text-shadow:0 2px 12px rgba(50,130,200,.35) !important; margin:0 0 .4rem !important;
}
.hero-header p { color:rgba(255,255,255,.95) !important; font-size:1.05rem !important; margin:0 !important; }
.hero-emoji { font-size:1.1rem; opacity:.75; letter-spacing:.5rem; margin-top:.6rem; }
.card {
background:rgba(255,255,255,.78) !important; backdrop-filter:blur(12px);
border-radius:20px !important; border:1.5px solid rgba(126,206,247,.5) !important;
box-shadow:0 4px 24px rgba(91,184,245,.15) !important; padding:1.2rem 1.2rem .8rem !important; margin-bottom:1rem !important;
}
/* Gradio gr.Group / gr.Column / gr.Row inner wrappers — make transparent so .card bg shows through */
.card > .form,
.card > div,
.card .block,
.card .gap,
.card .gr-group,
.card .gr-block,
.card .gr-box,
.card [data-node-id],
.card .svelte-1gfkn6j,
.card > .wrap {
background:transparent !important;
border:none !important;
box-shadow:none !important;
padding:0 !important;
}
.section-title { font-size:1rem !important; font-weight:700 !important; color:#3a87c8 !important; margin:0 0 .7rem !important; }
#run-btn {
background:linear-gradient(135deg,#4fb3f0 0%,#6ecbf7 50%,#9ddcff 100%) !important;
border:none !important; border-radius:50px !important; color:#fff !important;
font-size:1.15rem !important; font-weight:700 !important; letter-spacing:.04em;
padding:.85rem 1.5rem !important; box-shadow:0 6px 20px rgba(79,179,240,.45) !important;
transition:transform .18s,box-shadow .18s !important; width:100%;
}
#run-btn:hover { transform:translateY(-2px) scale(1.02) !important; box-shadow:0 10px 28px rgba(79,179,240,.6) !important; }
#run-btn:active { transform:translateY(0) scale(.98) !important; }
.gradio-accordion {
background:rgba(255,255,255,.65) !important; border-radius:16px !important;
border:1.5px solid rgba(126,206,247,.45) !important; margin-bottom:.7rem !important; overflow:hidden;
}
.gradio-accordion .label-wrap {
background:linear-gradient(90deg,#d6eeff 0%,#c2e5ff 100%) !important;
border-radius:14px 14px 0 0 !important; padding:.55rem 1rem !important;
font-weight:700 !important; color:#3a87c8 !important;
}
/* ── Gradio 4 theme variables: eliminate all grey ── */
:root,
.gradio-container {
--block-background-fill: #eaf6ff !important;
--block-border-color: #aad9f7 !important;
--block-border-width: 1.5px !important;
--block-label-background-fill: #d6eeff !important;
--block-label-text-color: #3a87c8 !important;
--block-shadow: 0 2px 8px rgba(91,184,245,.1) !important;
--input-background-fill: #f0f9ff !important;
--input-border-color: #aad9f7 !important;
--input-border-width: 1.5px !important;
--panel-background-fill: #eaf6ff !important;
--panel-border-color: #aad9f7 !important;
--background-fill-primary: #eaf6ff !important;
--background-fill-secondary: #d6eeff !important;
--border-color-primary: #aad9f7 !important;
--border-color-accent: #5bb8f5 !important;
--color-accent: #5bb8f5 !important;
--color-accent-soft: #d6eeff !important;
--neutral-50: #eaf6ff !important;
--neutral-100: #d6eeff !important;
--neutral-200: #aad9f7 !important;
}
/* ── Textbox actual input elements ── */
.gradio-textbox textarea,
.gradio-textbox input,
[data-testid="textbox"] textarea,
[data-testid="textbox"] input {
border-radius:12px !important;
border:1.5px solid #aad9f7 !important;
background:#f0f9ff !important;
font-family:inherit !important;
}
.gradio-textbox textarea:focus,
.gradio-textbox input:focus,
[data-testid="textbox"] textarea:focus,
[data-testid="textbox"] input:focus {
border-color:#5bb8f5 !important;
box-shadow:0 0 0 3px rgba(91,184,245,.2) !important;
background:#ffffff !important;
}
/* ── output-box / log-box overrides ── */
.output-box textarea { background:#edf7ff !important; border:1.5px solid #aad9f7 !important; border-radius:12px !important; }
.log-box textarea {
background:#e4f3ff !important; border:1.5px solid #93d4fb !important; border-radius:12px !important;
font-size:.82rem !important; font-family:'Courier New',monospace !important; white-space:pre-wrap !important;
}
input[type=range] { accent-color:#5bb8f5 !important; }
/* ── Audio widget: override all grey surfaces ── */
.gradio-audio,
.gradio-audio > div,
[data-testid="audio"],
[data-testid="audio"] > div,
[data-testid="audio"] > div > div {
border-radius:16px !important;
border:1.5px solid #aad9f7 !important;
background:#eaf6ff !important;
}
/* waveform / player bar */
.gradio-audio .waveform-container,
.gradio-audio audio,
[data-testid="audio"] .waveform-container,
[data-testid="audio"] audio {
background:#eaf6ff !important;
border-radius:12px !important;
}
/* toolbar row (mic / upload buttons area) */
.gradio-audio .controls,
.gradio-audio .icon-buttons,
.gradio-audio .recording-toolbar,
[data-testid="audio"] .controls,
[data-testid="audio"] .icon-buttons,
[data-testid="audio"] .recording-toolbar {
background:#d6eeff !important;
border-radius:0 0 14px 14px !important;
}
/* WaveSurfer canvas background */
.gradio-audio wave,
[data-testid="audio"] wave {
background:#eaf6ff !important;
}
/* grey block that wraps the whole upload/record area */
.gradio-audio .wrap,
[data-testid="audio"] .wrap,
.gradio-audio .empty,
[data-testid="audio"] .empty {
background:#eaf6ff !important;
border-color:#aad9f7 !important;
border-radius:14px !important;
}
/* label row above the widget */
.gradio-audio label span,
[data-testid="audio"] label span {
color:#3a87c8 !important;
font-weight:700 !important;
}
.step-badge {
display:inline-block; background:linear-gradient(90deg,#5bb8f5,#93d4fb); color:#fff;
font-size:.72rem; font-weight:700; padding:.18em .7em; border-radius:20px;
letter-spacing:.06em; margin-bottom:.3rem;
}
.output-box textarea {
background:linear-gradient(135deg,rgba(255,255,255,.92) 0%,rgba(214,238,255,.65) 100%) !important;
border:1.5px solid #aad9f7 !important; border-radius:12px !important;
}
.log-box textarea {
background:linear-gradient(135deg,rgba(210,236,255,.75) 0%,rgba(235,248,255,.6) 100%) !important;
border:1.5px solid #93d4fb !important; border-radius:12px !important;
font-size:.82rem !important; font-family:'Courier New',monospace !important;
white-space:pre-wrap !important;
}
footer { display:none !important; }
"""
with gr.Blocks(title="Everything becomes Nyan 🐱") as demo:
gr.HTML("""
""")
with gr.Row(equal_height=False):
with gr.Column(scale=5):
with gr.Group(elem_classes="card"):
gr.HTML('🎙️ Input Audio
')
audio_input = gr.Audio(
label="Record or upload an audio file",
sources=["microphone","upload"], type="filepath",
)
run_btn = gr.Button("🌟 Start Conversion!", variant="primary", size="lg", elem_id="run-btn")
with gr.Accordion("⚙️ Model Settings", open=False):
base_repo_in = gr.Textbox(label="Base Model (HuggingFace Hub ID)", value=BASE_REPO)
hub_id_in = gr.Textbox(label="Custom Model Hub ID (optional)", placeholder="your-username/your-model")
with gr.Accordion("🔊 TTS Voice Settings", open=True):
voice_in = gr.Dropdown(label="TTS Voice", choices=list(VOICE_PROMPTS.keys()), value=DEFAULT_VOICE)
with gr.Accordion("🚫 Toxic Word Settings", open=False):
with gr.Row():
extra_toxic_in = gr.Textbox(label="Additional Toxic Words (comma-separated)", placeholder="word1, word2")
replacement_in = gr.Textbox(label="Replacement String", value=DEFAULT_REPLACEMENT)
with gr.Accordion("🎛️ Generation Parameters", open=False):
asr_max_tokens_in = gr.Slider(label="ASR Max Tokens", minimum=64, maximum=1024, step=64, value=512)
tts_max_tokens_in = gr.Slider(label="TTS Max Tokens", minimum=64, maximum=2048, step=64, value=1024)
audio_temp_in = gr.Slider(label="Audio Temperature", minimum=0.1, maximum=2.0, step=0.1, value=1.0)
audio_topk_in = gr.Slider(label="Audio Top-K", minimum=1, maximum=128, step=1, value=4)
with gr.Column(scale=5):
with gr.Group(elem_classes="card"):
gr.HTML('🎵 Generated Audio
')
audio_output = gr.Audio(label="", type="filepath", interactive=False)
with gr.Group(elem_classes="card"):
gr.HTML('📝 Text Results
')
with gr.Row():
with gr.Column():
gr.HTML('STEP 1 — ASR')
transcript_out = gr.Textbox(
label="Recognized Text (before filter)", lines=4, interactive=False,
placeholder="Recognized text will appear here",
elem_classes="output-box",
)
with gr.Column():
gr.HTML('STEP 2 — Filtered')
filtered_out = gr.Textbox(
label="Filtered Text", lines=4, interactive=False,
placeholder="Filtered text will appear here",
elem_classes="output-box",
)
with gr.Group(elem_classes="card"):
gr.HTML('📋 Processing Log
')
status_out = gr.Textbox(
label="", lines=10, interactive=False,
placeholder="Logs will appear here after conversion starts…",
elem_classes="log-box",
)
with gr.Accordion("📋 Default Toxic Word List (click to expand)", open=False):
gr.Markdown("The following words are replaced with 'にゃん' by default.\n\n"
+ " ".join(f"`{w}`" for w in DEFAULT_TOXIC_WORDS))
run_btn.click(
fn=pipeline,
inputs=[
audio_input, hub_id_in, base_repo_in, voice_in,
extra_toxic_in, replacement_in,
asr_max_tokens_in, tts_max_tokens_in,
audio_temp_in, audio_topk_in,
],
outputs=[audio_output, transcript_out, filtered_out, status_out],
)
if __name__ == "__main__":
demo.queue().launch(
theme=gr.themes.Soft(
primary_hue="sky",
secondary_hue="blue",
neutral_hue="sky",
font=gr.themes.GoogleFont("Nunito"),
).set(
block_background_fill="#eaf6ff",
block_border_color="#aad9f7",
block_label_background_fill="#d6eeff",
input_background_fill="#f0f9ff",
panel_background_fill="#eaf6ff",
background_fill_primary="#eaf6ff",
background_fill_secondary="#d6eeff",
border_color_primary="#aad9f7",
),
css=CSS,
)