dublab / app.py
khantbhko's picture
DubLab Gradio app staged as Static Space (grant pending — sdk flips to gradio on approval)
92a6671 verified
Raw History Blame Contribute Delete
17.3 kB
"""DubLab on Hugging Face — Gradio ZeroGPU Space version.
Same pipeline as the Flask site (Spain), rebuilt as a Gradio app:
upload a short video (or paste a link) -> dubbed video back.
Differences from the Flask version:
- Gradio UI instead of Flask + custom HTML/JS (no admin panel, no
reviews, no blur-box picker — the core dub feature only).
- No @spaces.GPU anywhere: the whole pipeline is CPU-bound
(faster-whisper CPU, Nara ONNX CPU, ffmpeg, Gemini API), so it runs
on the Space's CPU and consumes ZERO daily GPU quota.
- Luna translation is skipped (no OpenCode agent on HF) — Gemini
translates directly, with the local NLLB fallback only if its model
is present (it is not bundled here).
- Limits are tighter because the Space CPU is small: 3 min / 100 MB
(the Spain site takes 10 min / 250 MB).
Secrets (Space Settings -> Variables and secrets):
GEMINI_API_KEY (required) — translation + Gemini TTS fallback
GEMINI_TTS_API_KEYS (optional) — comma-separated rotation keys
"""
import os
import re
import shutil
import subprocess
import sys
import time
import uuid
from pathlib import Path
import gradio as gr
# No OpenCode agent on HF: Luna is never imported here at all.
os.environ.setdefault("DUBLAB_NO_LUNA", "1")
os.environ.setdefault("NARA_THREADS", "2")
from alerts import quota_alert, quota_ok
from dub import (build_dub_track, cleanup_segments, configure_tts,
extract_audio, has_audio_stream, has_video_stream,
merge_segments, mux, probe_duration, probe_size,
set_engine, set_nara_voice, transcribe,
valid_nara_voice, write_ass, write_srt)
from gemini import translate_lines
from gemini_tts import QuotaExhausted
from linkutil import (MAX_DURATION, MAX_MB, check_link, download_link,
fetch_link_info)
BASE = Path(__file__).parent
JOBS = BASE / "jobs"
JOBS.mkdir(exist_ok=True)
TTL = 24 * 3600
VOICES = {
"en2mm": {"female": "my-MM-NilarNeural", "male": "my-MM-ThihaNeural"},
"mm2en": {"female": "en-US-AvaNeural", "male": "en-US-AndrewNeural"},
}
STAGE_LABEL = {
"queue": "တန်းစီနေတယ်",
"download": "ဗီဒီယို ဆွဲချနေတယ်",
"audio": "အသံထုတ်ယူနေတယ်",
"transcribe": "စကားပြောမှတ်နေတယ်",
"translate": "ဘာသာပြန်နေတယ်",
"voice": "အသံဖော်နေတယ်",
"mux": "ဗီဒီယိုပေါင်းနေတယ်",
"done": "ပြီးပြီ ✅",
}
# ---- startup checks -------------------------------------------------------
def _check_ffmpeg():
"""ffmpeg/ffprobe come from packages.txt. imageio-ffmpeg ships an
ffmpeg binary only (no ffprobe), so it is a fallback for ffmpeg
alone; without ffprobe the job cannot probe files and must fail
with a clear message instead of a cryptic one."""
if shutil.which("ffmpeg") and shutil.which("ffprobe"):
return
try:
import imageio_ffmpeg
exe = Path(imageio_ffmpeg.get_ffmpeg_exe())
os.environ["PATH"] = (str(exe.parent) + os.pathsep
+ os.environ.get("PATH", ""))
except Exception: # noqa: BLE001 - fall through to the loud check
pass
if not (shutil.which("ffmpeg") and shutil.which("ffprobe")):
raise RuntimeError("ffmpeg/ffprobe not found — packages.txt "
"did not install them; check the Space build log")
def ensure_nara():
"""Download the Nara Mini voices (Shoon + Ko Banyar, ~365 MB) from
the public model repo YeeNyunn/nara-mini-v1 on first start.
Gradio Spaces have no Dockerfile build step, so this runs at app
startup instead. The files land in ./nara-mini, the same layout
nara_tts.py expects (NARA_MODEL_DIR default).
"""
target = BASE / "nara-mini" / "onnx" / "shoon" / "encoder.onnx"
if target.exists():
print("[dublab] nara-mini voices already present")
return
print("[dublab] downloading Nara Mini voices (~365 MB)…")
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="YeeNyunn/nara-mini-v1", local_dir=str(BASE / "nara-mini"),
allow_patterns=["config.json", "frontend/*",
"onnx/shoon/*", "onnx/ko-banyar/*"])
print("[dublab] nara-mini voices ready")
def _sweep_old():
now = time.time()
for d in JOBS.iterdir():
try:
if d.is_dir() and now - d.stat().st_mtime > TTL:
shutil.rmtree(d, ignore_errors=True)
except OSError:
pass
_check_ffmpeg()
ensure_nara()
_sweep_old()
_tts_keys = [k.strip() for k in
os.environ.get("GEMINI_TTS_API_KEYS",
os.environ.get("GEMINI_API_KEY", "")).split(",")
if k.strip()]
configure_tts(_tts_keys)
KEY_OK = bool(os.environ.get("GEMINI_API_KEY"))
# ---- helpers --------------------------------------------------------------
def _is_quota(e):
if isinstance(e, QuotaExhausted):
return True
s = str(e).lower()
return "resource_exhausted" in s or "quota" in s or "429" in s
def clean_err(e):
"""User-facing error text — never leak server paths."""
s = str(e)
if ("speech.platform.bing.com" in s or "WSServerHandshakeError" in s
or "Invalid response status" in s):
return ("အသံထုတ်တဲ့ service (Microsoft) ခေတ္တ ပျက်နေတယ် — "
"ပြန်ကောင်းအောင် ပြင်နေပါတယ်။ ခဏနေမှ ပြန်စမ်းပါ 🙏")
return s.replace(str(BASE), "[app]").replace(str(JOBS), "[jobs]")[:300]
# ---- the dub job (generator: yields status updates as it runs) -----------
def dub_video(video_file, link, direction, voice, engine, subs, sub_bg,
progress=gr.Progress(track_tqdm=True)):
log = []
def say(stage, msg=""):
label = STAGE_LABEL.get(stage, stage)
line = f"[{time.strftime('%H:%M:%S')}] {label}" + (f" — {msg}" if msg else "")
log.append(line)
return line
def out(status, video=None, srt=None):
return status, video, srt, "\n".join(log)
jid = uuid.uuid4().hex[:12]
d = JOBS / jid
d.mkdir(parents=True)
try:
if not KEY_OK:
raise RuntimeError(
"server key မရှိသေးဘူး — Space owner က GEMINI_API_KEY ကို "
"Space Settings > Secrets မှာ ထည့်ရမယ်")
direction = direction or "en2mm"
voice = voice or "female"
engine = engine or "nara"
if engine not in ("edge", "gemini", "nara"):
engine = "nara"
nara_voice = valid_nara_voice("shoon" if voice == "female" else "ko-banyar")
api_key = os.environ.get("GEMINI_API_KEY")
# ---- input: pasted link wins over an uploaded file ----
url = (link or "").strip()
ext = ".mp4"
if url:
say("download", "link စစ်နေတယ်…")
yield out(say("download", "link စစ်နေတယ်…"))
url = check_link(url)
fetch_link_info(url) # fail fast on bad/private/too-long links
say("download", "ဗီဒီယို ဆွဲချနေတယ်…")
yield out(say("download", "ဗီဒီယို ဆွဲချနေတယ်…"))
ext = download_link(url, d)
else:
if not video_file:
raise ValueError("ဗီဒီယိုဖိုင် ရွေးပေးပါ — ဒါမှမဟုတ် link paste လုပ်ပါ")
src_in = Path(video_file)
ext = src_in.suffix.lower()
if ext not in (".mp4", ".mov", ".mkv", ".webm", ".m4v"):
raise ValueError("MP4/MOV ဖိုင်ပဲ လက်ခံတယ်")
if src_in.stat().st_size > MAX_MB * 1024 * 1024:
raise ValueError(f"{MAX_MB} MB ထက် ကြီးတဲ့ ဖိုင် လက်မခံဘူး")
shutil.copyfile(src_in, d / ("input" + ext))
src = d / ("input" + ext)
total = probe_duration(src)
if total > MAX_DURATION:
raise ValueError("၃ မိနစ်ထက် ရှည်တဲ့ ဗီဒီယို လက်မခံဘူး (Space version)")
if not has_video_stream(src):
raise ValueError("ဗီဒီယိုဖိုင် မဟုတ်ဘူး")
if not has_audio_stream(src):
raise ValueError("ဒီဗီဒီယိုမှာ အသံမပါဘူး — အသံပါတဲ့ ဗီဒီယိုမှ dub လို့ရမယ်")
say("audio", f"video {total:.1f}s")
progress(0.05, desc="အသံထုတ်ယူနေတယ်…")
yield out(say("audio", "အသံထုတ်ယူနေတယ်…"))
wav = d / "audio.wav"
extract_audio(src, wav)
progress(0.08, desc="စကားပြောမှတ်နေတယ်…")
yield out(say("transcribe", "Whisper နဲ့ စကားပြောမှတ်နေတယ်…"))
lang_hint = "en" if direction == "en2mm" else "my"
segs = transcribe(wav, lang_hint)
if not segs:
raise ValueError("အသံထဲမှာ စကားပြောသံ မတွေ့ဘူး")
utter = merge_segments(segs)
say("transcribe", f"{len(segs)} segments → {len(utter)} utterances")
progress(0.45, desc="ဘာသာပြန်နေတယ်…")
yield out(say("translate", "Gemini နဲ့ ဘာသာပြန်နေတယ်…"))
try:
texts = translate_lines([u["text"] for u in utter], direction, api_key)
say("translate", "Gemini နဲ့ ပြီးပြီ")
except Exception as e2: # noqa: BLE001 - Gemini down/quota: local fallback
say("translate", f"Gemini ခဏမရလို့ ({clean_err(e2)}) — NLLB (local) နဲ့ ဆက်ဘာသာပြန်တယ်")
try:
from nllb_translate import translate_lines as nllb_translate_lines
texts = nllb_translate_lines([u["text"] for u in utter], direction)
say("translate", "NLLB (local) နဲ့ ပြီးပြီ")
except Exception:
# NLLB model is not bundled on HF: surface the Gemini
# error instead of masking it with a missing-model one.
raise e2
for u, t in zip(utter, texts):
u["text"] = t or u["text"]
voice_name = VOICES[direction][voice]
set_engine(engine)
set_nara_voice(nara_voice)
progress(0.58, desc="အသံဖော်နေတယ်…")
yield out(say("voice", f"{len(utter)} လိုင်း အသံဖော်နေတယ် ({voice_name})…"))
def tts_prog(i, n):
progress(0.58 + 0.32 * i / max(n, 1), desc=f"အသံဖော်နေတယ်… {i}/{n}")
dub_wav, warp, cues = build_dub_track(utter, voice_name, d, total, tts_prog)
slowed = [f for (_a, _b, f) in warp if f > 1.001]
if slowed:
say("voice", f"စကားပြောသံ သဘာဝကျအောင် scene {len(slowed)} ခု နှေးထားတယ်")
progress(0.93, desc="ဗီဒီယိုပေါင်းနေတယ်…")
yield out(say("mux", "ဗီဒီယိုပေါင်းနေတယ်…"))
ass_path = None
srt_path = None
if subs:
vw, vh = probe_size(src)
write_ass(cues, d / "subs.ass", vw, vh, sub_bg or "box")
write_srt(cues, d / "subs.srt")
ass_path = d / "subs.ass"
srt_path = str(d / "subs.srt")
say("mux", f"စာတန်း {len(cues)} လိုင်း ထည့်နေတယ်…")
mux(src, dub_wav, d / "dubbed.mp4", warp, ass_path)
try:
cleanup_segments(d)
except OSError:
pass
progress(1.0, desc="ပြီးပြီ ✅")
say("done", "ပြီးပြီ ✅ — အောက်က video ကို ကြည့်နိုင် / download ဆွဲနိုင်ပါတယ်")
quota_ok()
yield out("✅ **ပြီးပြီ!** — အောက်က video ကို ကြည့်နိုင် / download ဆွဲနိုင်ပါတယ်",
str(d / "dubbed.mp4"), srt_path)
except Exception as e: # noqa: BLE001 - surfaced to UI
err = clean_err(e)
log.append(f"[{time.strftime('%H:%M:%S')}] FAILED: {type(e).__name__}: {err}")
if _is_quota(e):
quota_alert("gradio space", err)
yield out(f"❌ **{err}**", None, None)
finally:
# keep the result files; the job dir itself is swept after a day
pass
# ---- UI -------------------------------------------------------------------
KEY_BANNER = "" if KEY_OK else (
"\n\n> ⚠️ **Server key မရှိသေးဘူး** — Space owner က `GEMINI_API_KEY` ကို "
"Space Settings → Secrets မှာ ထည့်ရမယ်။ Key မရှိရင် dub job တွေ အလုပ်မလုပ်ဘူး။\n")
with gr.Blocks(title="DubLab — EN↔MM Video Dubbing") as demo:
gr.Markdown(
"# 🎬 DubLab — English ↔ Burmese Video Dubbing\n"
"ဗီဒီယိုအတိုတင်လိုက် → အခြားဘာသာစကားနဲ့ dub လုပ်ပြီးသား video ပြန်ရမယ်။\n"
"Free Hugging Face Space — Whisper (local) → Gemini translate → "
"Nara Mini အသံ (Shoon / Ko Banyar, local, free).\n\n"
"**Limits (Space version):** video ≤ 3 min, ≤ 100 MB. "
"ပိုရှည်တဲ့ video တွေအတွက် [Spain site](https://dublab.kbko.qzz.io) "
"(10 min / 250 MB) သုံးပါ." + KEY_BANNER)
with gr.Row():
with gr.Column():
# Gradio 6: gr.Video has no `type` param — an uploaded video
# arrives as a filepath string, which is what we want.
video_in = gr.Video(label="🎥 Video file (MP4 / MOV / MKV / WebM)")
link_in = gr.Textbox(label="🔗 ဒါမှမဟုတ် video link (YouTube / TikTok / RedNote / Instagram / Facebook / X / Vimeo / Bilibili)",
placeholder="https://… — link ထည့်ရင် file မလိုဘူး")
direction = gr.Radio(["en2mm", "mm2en"], value="en2mm",
label="ဘာသာပြန်လမ်း",
info="en2mm = English video → မြန်မာအသံ · mm2en = မြန်မာစကား video → English အသံ")
voice = gr.Radio(["female", "male"], value="female",
label="အသံ",
info="female = Shoon · male = Ko Banyar (Nara Mini local voices)")
engine = gr.Radio(["nara", "edge", "gemini"], value="nara",
label="အသံထုတ်စက် (engine)",
info="nara = local, free (default) · edge = Microsoft · gemini = Gemini TTS")
with gr.Row():
subs = gr.Checkbox(label="စာတန်း burn-in လုပ်မလား", value=False)
sub_bg = gr.Radio(["box", "none"], value="box", label="စာတန်း နောက်ခံ")
submit = gr.Button("🎙️ Dub လုပ်ပါ", variant="primary")
with gr.Column():
status = gr.Markdown("အသင့်ဖြစ်ပြီ — video တင် (သို့) link paste လုပ်ပြီး **Dub လုပ်ပါ** ကို နှိပ်ပါ။")
video_out = gr.Video(label="✅ Dub လုပ်ပြီးသား video", autoplay=False)
srt_out = gr.File(label="📄 Subtitle (.srt) — စာတန်း burn-in ရွေးထားရင် ရမယ်")
log_box = gr.Textbox(label="Log", lines=10, interactive=False, autoscroll=True)
submit.click(
dub_video,
inputs=[video_in, link_in, direction, voice, engine, subs, sub_bg],
outputs=[status, video_out, srt_out, log_box],
)
gr.Markdown(
"---\n"
"DubLab by [Khant Bhone Ko](https://pro.kxant.org) · MIT license · "
"[Source (Docker version)](https://huggingface.co/khantbhko/dublab) · "
"[GitHub](https://github.com/khantbhoneko36-commits/dublab) · "
"Made with ❤️ for Myanmar creators")
demo.queue(max_size=8, default_concurrency_limit=1)
if __name__ == "__main__":
# Explicit bind: HF Spaces serve on $PORT (7860); local runs use 7860 too.
demo.launch(server_name="0.0.0.0",
server_port=int(os.environ.get("PORT", "7860")))