Files
whisper-stt/app/tasks.py

415 lines
20 KiB
Python

"""
STT + Subtitle Pipeline Celery Tasks
subtitle_pipeline_task:
Step 1: ffmpeg → 16kHz WAV 추출
Step 2: Whisper → 원어 SRT / VTT 생성
Step 3: LLM → 번역 SRT / VTT 생성 (선택)
"""
import os, json, subprocess, tempfile
import httpx
from celery import Celery
from ocr_tasks import ocr_task # noqa: F401
REDIS_URL = os.getenv("REDIS_URL", "redis://redis:6379/0")
MODEL_SIZE = os.getenv("WHISPER_MODEL", "medium")
DEVICE = os.getenv("WHISPER_DEVICE", "cpu")
COMPUTE_TYPE = os.getenv("WHISPER_COMPUTE_TYPE", "int8")
LANGUAGE = os.getenv("WHISPER_LANGUAGE", "ko") or None
BEAM_SIZE = int(os.getenv("WHISPER_BEAM_SIZE", "5"))
INITIAL_PROMPT = os.getenv("WHISPER_INITIAL_PROMPT", "") or None
OUTPUT_DIR = os.getenv("OUTPUT_DIR", "/data/outputs")
OLLAMA_URL = os.getenv("OLLAMA_URL", "http://192.168.0.126:11434")
OLLAMA_TIMEOUT = int(os.getenv("OLLAMA_TIMEOUT", "600"))
_cpu_threads_env = int(os.getenv("CPU_THREADS", "0"))
CPU_THREADS = _cpu_threads_env if _cpu_threads_env > 0 else None
celery_app = Celery("whisper_tasks", broker=REDIS_URL, backend=REDIS_URL)
celery_app.conf.update(
task_serializer="json", result_serializer="json",
accept_content=["json"], task_track_started=True, result_expires=3600,
)
_whisper_model = None
def get_model():
global _whisper_model
if _whisper_model is None:
from faster_whisper import WhisperModel
kwargs = dict(device=DEVICE, compute_type=COMPUTE_TYPE)
if CPU_THREADS is not None: kwargs["cpu_threads"] = CPU_THREADS
print(f"[Whisper] 로딩: {MODEL_SIZE}/{DEVICE}/{COMPUTE_TYPE}/threads={CPU_THREADS or 'auto'}")
_whisper_model = WhisperModel(MODEL_SIZE, **kwargs)
print("[Whisper] 로드 완료")
return _whisper_model
# ══════════════════════════════════════════════════════════════
# 언어 코드 → 표시명
# ══════════════════════════════════════════════════════════════
LANG_NAMES = {
"ko":"한국어","en":"English","ja":"日本語","zh":"中文(简体)",
"zh-tw":"中文(繁體)","fr":"Français","de":"Deutsch","es":"Español",
"it":"Italiano","pt":"Português","ru":"Русский","ar":"العربية",
"vi":"Tiếng Việt","th":"ไทย","id":"Bahasa Indonesia",
"nl":"Nederlands","pl":"Polski","tr":"Türkçe","sv":"Svenska",
"uk":"Українська","hi":"हिन्दी","bn":"বাংলা",
}
def _lang_name(code): return LANG_NAMES.get(code, code)
# ══════════════════════════════════════════════════════════════
# 자막 포맷 생성
# ══════════════════════════════════════════════════════════════
def _srt_time(s: float) -> str:
ms = int(round(s * 1000))
h, r = divmod(ms, 3600000); m, r = divmod(r, 60000); sec, ms = divmod(r, 1000)
return f"{h:02d}:{m:02d}:{sec:02d},{ms:03d}"
def _vtt_time(s: float) -> str:
return _srt_time(s).replace(",", ".")
def make_srt(segments: list) -> str:
out = []
for i, seg in enumerate(segments, 1):
out += [str(i), f"{_srt_time(seg['start'])} --> {_srt_time(seg['end'])}", seg["text"].strip(), ""]
return "\n".join(out)
def make_vtt(segments: list) -> str:
out = ["WEBVTT", ""]
for i, seg in enumerate(segments, 1):
out += [str(i), f"{_vtt_time(seg['start'])} --> {_vtt_time(seg['end'])}", seg["text"].strip(), ""]
return "\n".join(out)
# ══════════════════════════════════════════════════════════════
# LLM 번역 (세그먼트 배치)
# ══════════════════════════════════════════════════════════════
def _translate_batch(texts: list, target_lang: str,
use_openrouter: bool, model: str,
openrouter_url: str, openrouter_key: str) -> list:
"""texts 리스트 → 번역된 texts 리스트"""
if not texts or not model: return texts
lang_name = _lang_name(target_lang)
prompt = (
f"아래 자막 문장 배열을 {lang_name}로 번역해줘.\n"
f"반드시 JSON 문자열 배열로만 답해. 설명·마크다운 없이 배열만 출력.\n"
f"입력과 동일한 개수와 순서를 유지해.\n\n"
f"{json.dumps(texts, ensure_ascii=False)}"
)
try:
if use_openrouter and openrouter_key:
resp = httpx.post(
f"{openrouter_url.rstrip('/')}/chat/completions",
headers={"Authorization": f"Bearer {openrouter_key}",
"HTTP-Referer": "https://voicescript.local",
"Content-Type": "application/json"},
json={"model": model,
"messages": [{"role":"user","content":prompt}],
"temperature": 0.2},
timeout=float(OLLAMA_TIMEOUT),
)
resp.raise_for_status()
raw = resp.json()["choices"][0]["message"]["content"].strip()
else:
resp = httpx.post(f"{OLLAMA_URL}/api/chat",
json={"model": model,
"messages": [{"role":"user","content":prompt}],
"stream": False, "options": {"temperature": 0.2}},
timeout=float(OLLAMA_TIMEOUT))
resp.raise_for_status()
raw = resp.json().get("message",{}).get("content","").strip()
# 코드블록 제거 후 JSON 파싱
if "```" in raw:
raw = raw.split("```")[1].lstrip("json\n").rstrip()
result = json.loads(raw)
if isinstance(result, list) and len(result) == len(texts):
return [str(r) for r in result]
return texts
except Exception as e:
print(f"[번역 실패] {e}")
return texts # 실패 시 원본 유지
# ══════════════════════════════════════════════════════════════
# STT + Ollama/OpenRouter 후처리 (기존 음성변환용)
# ══════════════════════════════════════════════════════════════
def _ollama_postprocess(text: str, model: str) -> str:
if not model or not text.strip(): return text
prompt = ("다음은 음성 인식으로 추출된 텍스트입니다. "
"내용은 절대 변경하지 말고, 문장 부호를 추가하고 자연스럽게 다듬어줘. "
"결과 텍스트만 출력하고 설명은 하지 마.\n\n" + text)
try:
resp = httpx.post(f"{OLLAMA_URL}/api/chat",
json={"model":model,"messages":[{"role":"user","content":prompt}],
"stream":False,"options":{"temperature":0.1}},
timeout=float(OLLAMA_TIMEOUT))
resp.raise_for_status()
return resp.json().get("message",{}).get("content","").strip() or text
except: return text
def _openrouter_postprocess(text: str, model: str, base_url: str, api_key: str) -> str:
if not model or not api_key or not text.strip(): return text
prompt = ("다음은 음성 인식으로 추출된 텍스트입니다. "
"내용은 절대 변경하지 말고, 문장 부호를 추가하고 자연스럽게 다듬어줘. "
"결과 텍스트만 출력하고 설명은 하지 마.\n\n" + text)
try:
resp = httpx.post(f"{base_url.rstrip('/')}/chat/completions",
headers={"Authorization":f"Bearer {api_key}","HTTP-Referer":"https://voicescript.local","Content-Type":"application/json"},
json={"model":model,"messages":[{"role":"user","content":prompt}],"temperature":0.1},
timeout=float(OLLAMA_TIMEOUT))
resp.raise_for_status()
return resp.json()["choices"][0]["message"]["content"].strip() or text
except: return text
# ══════════════════════════════════════════════════════════════
# 기존 STT 태스크 (음성변환 탭용)
# ══════════════════════════════════════════════════════════════
@celery_app.task(bind=True, name="tasks.transcribe_task", queue="stt")
def transcribe_task(
self,
file_id: str, audio_path: str,
use_ollama: bool = False, ollama_model: str = "",
use_openrouter: bool = False, openrouter_model: str = "",
openrouter_url: str = "", openrouter_key: str = "",
):
self.update_state(state="PROGRESS", meta={"progress":5,"message":"모델 준비 중..."})
try:
model = get_model()
self.update_state(state="PROGRESS", meta={"progress":15,"message":"오디오 분석 중..."})
segments_gen, info = model.transcribe(
audio_path, language=LANGUAGE, beam_size=BEAM_SIZE,
initial_prompt=INITIAL_PROMPT, vad_filter=True,
vad_parameters=dict(min_silence_duration_ms=500), word_timestamps=False,
)
self.update_state(state="PROGRESS", meta={"progress":30,"message":"텍스트 변환 중..."})
segments, parts = [], []
duration = info.duration
for seg in segments_gen:
segments.append({"start":round(seg.start,3),"end":round(seg.end,3),"text":seg.text.strip()})
parts.append(seg.text.strip())
if duration > 0:
pct = 30 + int((seg.end/duration)*50)
self.update_state(state="PROGRESS",
meta={"progress":min(pct,80),"message":f"변환 중... {seg.end:.0f}s / {duration:.0f}s"})
raw_text = "\n".join(parts)
full_text = raw_text
if use_ollama and ollama_model:
self.update_state(state="PROGRESS",meta={"progress":85,"message":f"Ollama({ollama_model}) 교정 중..."})
full_text = _ollama_postprocess(raw_text, ollama_model)
elif use_openrouter and openrouter_model and openrouter_key:
self.update_state(state="PROGRESS",meta={"progress":85,"message":f"OpenRouter({openrouter_model}) 교정 중..."})
full_text = _openrouter_postprocess(raw_text, openrouter_model, openrouter_url, openrouter_key)
self.update_state(state="PROGRESS",meta={"progress":95,"message":"파일 저장 중..."})
os.makedirs(OUTPUT_DIR, exist_ok=True)
output_filename = f"{file_id}.txt"
with open(os.path.join(OUTPUT_DIR, output_filename),"w",encoding="utf-8") as f:
f.write(f"# 변환 결과\n# 언어: {info.language} | 재생 시간: {duration:.1f}\n\n## 전체 텍스트\n\n{full_text}\n\n## 타임스탬프별 세그먼트\n\n")
for seg in segments:
m,s=divmod(int(seg['start']),60)
f.write(f"[{m:02d}:{s:02d}] {seg['text']}\n")
try: os.remove(audio_path)
except: pass
return {
"text":full_text,"raw_text":raw_text,"segments":segments,
"language":info.language,"duration":round(duration,1),
"output_file":output_filename,
"ollama_used":use_ollama and bool(ollama_model),
"ollama_model":ollama_model if (use_ollama and ollama_model) else "",
"openrouter_used":use_openrouter and bool(openrouter_model) and bool(openrouter_key),
"openrouter_model":openrouter_model if (use_openrouter and openrouter_model) else "",
}
except Exception as e:
raise Exception(f"변환 실패: {str(e)}")
# ══════════════════════════════════════════════════════════════
# 자막 파이프라인 태스크
# Step 1: ffmpeg → WAV
# Step 2: Whisper → 원어 SRT/VTT
# Step 3: LLM → 번역 SRT/VTT (선택)
# ══════════════════════════════════════════════════════════════
@celery_app.task(bind=True, name="tasks.subtitle_pipeline_task", queue="stt")
def subtitle_pipeline_task(
self,
file_id: str,
video_path: str,
src_language: str = "", # 원어 코드 (빈칸=자동)
subtitle_fmt: str = "srt", # srt | vtt | both
translate_to: str = "", # 번역 대상 (빈칸=번역 안 함)
trans_model: str = "", # 번역 모델
trans_via: str = "ollama",# ollama | openrouter
openrouter_url: str = "",
openrouter_key: str = "",
):
os.makedirs(OUTPUT_DIR, exist_ok=True)
wav_path = os.path.join(os.path.dirname(video_path), f"{file_id}_audio.wav")
result_files = {}
try:
# ── Step 1: ffmpeg 오디오 추출 ────────────────────────
self.update_state(state="PROGRESS", meta={
"progress": 5,
"step": 1,
"step_msg": "오디오 추출 중...",
"message": "Step 1/3 — ffmpeg 오디오 추출 중..."
})
cmd = [
"ffmpeg", "-y",
"-i", video_path,
"-vn", # 비디오 스트림 제거
"-ar", "16000", # 16kHz — Whisper 최적
"-ac", "1", # 모노
"-c:a", "pcm_s16le",# WAV 무손실
wav_path
]
proc = subprocess.run(cmd, capture_output=True, timeout=600)
if proc.returncode != 0:
err = proc.stderr.decode(errors="replace")[-500:]
raise Exception(f"ffmpeg 오디오 추출 실패: {err}")
if not os.path.exists(wav_path) or os.path.getsize(wav_path) < 1000:
raise Exception("ffmpeg가 오디오를 추출하지 못했습니다. 영상에 오디오 트랙이 있는지 확인하세요.")
try: os.remove(video_path)
except: pass
# ── Step 2: Whisper STT → 원어 자막 ───────────────────
self.update_state(state="PROGRESS", meta={
"progress": 15,
"step": 2,
"step_msg": "음성 인식 중...",
"message": "Step 2/3 — Whisper 음성 인식 시작..."
})
whisper = get_model()
lang = src_language.strip() or None
segments_gen, info = whisper.transcribe(
wav_path,
language=lang,
beam_size=BEAM_SIZE,
initial_prompt=INITIAL_PROMPT,
vad_filter=True,
vad_parameters=dict(min_silence_duration_ms=500),
word_timestamps=False,
)
segments = []
duration = info.duration
detected_lang = info.language
for seg in segments_gen:
segments.append({
"start": round(seg.start, 3),
"end": round(seg.end, 3),
"text": seg.text.strip(),
})
if duration > 0:
pct = 15 + int((seg.end / duration) * 55)
self.update_state(state="PROGRESS", meta={
"progress": min(pct, 70),
"step": 2,
"step_msg": f"{seg.end:.0f}s / {duration:.0f}s 인식 완료",
"message": f"Step 2/3 — {seg.end:.0f}s / {duration:.0f}s",
})
try: os.remove(wav_path)
except: pass
if not segments:
raise Exception("음성이 감지되지 않았습니다. 영상에 음성이 있는지 확인하세요.")
# 원어 자막 저장
lang_suffix = detected_lang
if subtitle_fmt in ("srt", "both"):
fn = f"{file_id}.{lang_suffix}.srt"
with open(os.path.join(OUTPUT_DIR, fn), "w", encoding="utf-8") as f:
f.write(make_srt(segments))
result_files["srt_orig"] = fn
if subtitle_fmt in ("vtt", "both"):
fn = f"{file_id}.{lang_suffix}.vtt"
with open(os.path.join(OUTPUT_DIR, fn), "w", encoding="utf-8") as f:
f.write(make_vtt(segments))
result_files["vtt_orig"] = fn
# ── Step 3: LLM 번역 (선택) ───────────────────────────
translated_segments = None
if translate_to and translate_to != detected_lang and trans_model:
target_name = _lang_name(translate_to)
use_or = (trans_via == "openrouter" and bool(openrouter_key))
total = len(segments)
CHUNK = 25 # 한 번에 25개씩 번역
translated_texts = []
for ci, start in enumerate(range(0, total, CHUNK)):
chunk = segments[start:start+CHUNK]
pct = 72 + int((ci * CHUNK / total) * 22)
self.update_state(state="PROGRESS", meta={
"progress": min(pct, 94),
"step": 3,
"step_msg": f"{min(start+CHUNK, total)}/{total}개 번역 완료",
"message": f"Step 3/3 — {target_name}로 번역 중... ({min(start+CHUNK,total)}/{total})",
})
batch_texts = [s["text"] for s in chunk]
translated = _translate_batch(
batch_texts, translate_to,
use_openrouter=use_or,
model=trans_model,
openrouter_url=openrouter_url,
openrouter_key=openrouter_key,
)
translated_texts.extend(translated)
# 번역된 텍스트 → 세그먼트 조합 (타임스탬프 유지)
translated_segments = [
{**seg, "text": translated_texts[i] if i < len(translated_texts) else seg["text"]}
for i, seg in enumerate(segments)
]
# 번역 자막 저장
trans_suffix = translate_to
if subtitle_fmt in ("srt", "both"):
fn = f"{file_id}.{trans_suffix}.srt"
with open(os.path.join(OUTPUT_DIR, fn), "w", encoding="utf-8") as f:
f.write(make_srt(translated_segments))
result_files["srt_trans"] = fn
if subtitle_fmt in ("vtt", "both"):
fn = f"{file_id}.{trans_suffix}.vtt"
with open(os.path.join(OUTPUT_DIR, fn), "w", encoding="utf-8") as f:
f.write(make_vtt(translated_segments))
result_files["vtt_trans"] = fn
self.update_state(state="PROGRESS", meta={
"progress": 98, "step": 3,
"step_msg": "완료", "message": "자막 파일 저장 완료"
})
return {
"detected_language": detected_lang,
"duration": round(duration, 1),
"segment_count": len(segments),
"translated": bool(translated_segments),
"translate_to": translate_to if translated_segments else "",
"subtitle_fmt": subtitle_fmt,
# 파일
"srt_orig": result_files.get("srt_orig", ""),
"vtt_orig": result_files.get("vtt_orig", ""),
"srt_trans": result_files.get("srt_trans", ""),
"vtt_trans": result_files.get("vtt_trans", ""),
}
except Exception as e:
# 임시 파일 정리
for p in [video_path, wav_path]:
try: os.remove(p)
except: pass
raise Exception(f"자막 생성 실패: {str(e)}")