Files
whisper-stt/app/tasks.py

383 lines
18 KiB
Python

"""
STT Celery Tasks
- faster-whisper 변환
- Ollama / OpenRouter 후처리 (교정 또는 번역)
- SRT / VTT / TXT 자막 파일 생성
"""
import os, json
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,
)
_model = None
def get_model():
global _model
if _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'}")
_model = WhisperModel(MODEL_SIZE, **kwargs)
print("[Whisper] 로드 완료")
return _model
# ══════════════════════════════════════════════════════════════
# 언어 코드 매핑
# ══════════════════════════════════════════════════════════════
LANG_NAMES = {
"ko":"한국어","en":"English","ja":"日本語","zh":"中文","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",
}
def _lang_name(code: str) -> str:
return LANG_NAMES.get(code, code)
# ══════════════════════════════════════════════════════════════
# 자막 포맷 생성
# ══════════════════════════════════════════════════════════════
def _fmt_srt_time(s: float) -> str:
"""초 → SRT 시간 포맷 00:00:00,000"""
ms = int(round(s * 1000))
h, rem = divmod(ms, 3600000)
m, rem = divmod(rem, 60000)
sec, ms = divmod(rem, 1000)
return f"{h:02d}:{m:02d}:{sec:02d},{ms:03d}"
def _fmt_vtt_time(s: float) -> str:
"""초 → VTT 시간 포맷 00:00:00.000"""
return _fmt_srt_time(s).replace(",", ".")
def _make_srt(segments: list) -> str:
lines = []
for i, seg in enumerate(segments, 1):
lines.append(str(i))
lines.append(f"{_fmt_srt_time(seg['start'])} --> {_fmt_srt_time(seg['end'])}")
lines.append(seg["text"].strip())
lines.append("")
return "\n".join(lines)
def _make_vtt(segments: list) -> str:
lines = ["WEBVTT", ""]
for i, seg in enumerate(segments, 1):
lines.append(f"{i}")
lines.append(f"{_fmt_vtt_time(seg['start'])} --> {_fmt_vtt_time(seg['end'])}")
lines.append(seg["text"].strip())
lines.append("")
return "\n".join(lines)
# ══════════════════════════════════════════════════════════════
# 번역 (Ollama / OpenRouter)
# ══════════════════════════════════════════════════════════════
def _translate_segments(segments: list, target_lang: str,
use_openrouter: bool, model: str,
openrouter_url: str, openrouter_key: str,
task_self=None) -> list:
"""세그먼트 텍스트를 target_lang으로 번역해서 새 세그먼트 리스트 반환"""
if not model or not target_lang:
return segments
lang_name = _lang_name(target_lang)
translated = []
# 세그먼트를 청크로 묶어서 번역 (API 호출 최소화)
# 최대 20개씩 묶음
CHUNK = 20
chunks = [segments[i:i+CHUNK] for i in range(0, len(segments), CHUNK)]
for ci, chunk in enumerate(chunks):
if task_self:
pct = 85 + int((ci / len(chunks)) * 10)
task_self.update_state(state="PROGRESS",
meta={"progress": pct,
"message": f"번역 중... ({ci*CHUNK+1}/{len(segments)})"})
# JSON 배열로 텍스트만 전달
texts = [seg["text"].strip() for seg in chunk]
prompt = (
f"다음 문장들을 {lang_name}로 번역해줘.\n"
f"JSON 배열 형식으로만 답해줘. 설명 없이 번역된 문장 배열만 출력해.\n"
f"입력 배열과 동일한 개수, 동일한 순서로 출력해.\n\n"
f"입력: {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()
# JSON 파싱
# 코드블록 제거
if "```" in raw:
raw = raw.split("```")[1].lstrip("json").strip()
trans_texts = json.loads(raw)
if not isinstance(trans_texts, list):
trans_texts = texts # 파싱 실패 시 원본 유지
except Exception as e:
print(f"[번역 실패 chunk {ci}] {e}")
trans_texts = texts # 실패 시 원본 유지
# 번역된 텍스트를 세그먼트에 결합
for seg, t_text in zip(chunk, trans_texts):
translated.append({**seg, "text": t_text})
# 남은 세그먼트 (번역 누락)
if len(trans_texts) < len(chunk):
for seg in chunk[len(trans_texts):]:
translated.append(seg)
return translated
# ══════════════════════════════════════════════════════════════
# Ollama 텍스트 후처리 (교정)
# ══════════════════════════════════════════════════════════════
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()
result = resp.json().get("message",{}).get("content","").strip()
return result if result else text
except Exception as e:
print(f"[Ollama 후처리 실패] {e}"); 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()
result = resp.json()["choices"][0]["message"]["content"].strip()
return result if result else text
except Exception as e:
print(f"[OpenRouter 후처리 실패] {e}"); return text
# ══════════════════════════════════════════════════════════════
# 메인 STT Task
# ══════════════════════════════════════════════════════════════
@celery_app.task(bind=True, name="tasks.transcribe_task", queue="stt")
def transcribe_task(
self,
file_id: str,
audio_path: str,
# 후처리
use_ollama: bool = False,
ollama_model: str = "",
use_openrouter: bool = False,
openrouter_model: str = "",
openrouter_url: str = "",
openrouter_key: str = "",
# 자막
subtitle_mode: bool = False, # True → 자막 파일 생성
subtitle_format: str = "srt", # srt | vtt | both
translate_to: str = "", # 번역 대상 언어 코드 (없으면 원어 자막)
translate_model: str = "", # 번역에 쓸 모델
translate_via: str = "ollama",# ollama | openrouter
# 원본 언어 강제 지정 (없으면 auto)
force_language: 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..."})
lang = force_language.strip() or LANGUAGE
segments_gen, info = model.transcribe(
audio_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,
)
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)*45)
self.update_state(state="PROGRESS",
meta={"progress":min(pct,75),
"message":f"변환 중... {seg.end:.0f}s / {duration:.0f}s"})
raw_text = "\n".join(parts)
full_text = raw_text
detected_lang = info.language
# ── 텍스트 후처리 (교정) ──────────────────────────────
if use_ollama and ollama_model and not subtitle_mode:
self.update_state(state="PROGRESS",
meta={"progress":80,"message":f"Ollama({ollama_model}) 교정 중..."})
full_text = _ollama_postprocess(raw_text, ollama_model)
elif use_openrouter and openrouter_model and openrouter_key and not subtitle_mode:
self.update_state(state="PROGRESS",
meta={"progress":80,"message":f"OpenRouter({openrouter_model}) 교정 중..."})
full_text = _openrouter_postprocess(raw_text, openrouter_model, openrouter_url, openrouter_key)
# ── 자막 모드: 번역 ──────────────────────────────────
translated_segments = segments
is_translated = False
if subtitle_mode and translate_to and translate_to != detected_lang:
t_model = translate_model or (ollama_model if translate_via=="ollama" else openrouter_model)
t_via_or = (translate_via == "openrouter" and bool(openrouter_key))
self.update_state(state="PROGRESS",
meta={"progress":82,
"message":f"{_lang_name(translate_to)}로 번역 중..."})
translated_segments = _translate_segments(
segments, translate_to,
use_openrouter=t_via_or,
model=t_model,
openrouter_url=openrouter_url,
openrouter_key=openrouter_key,
task_self=self,
)
is_translated = True
# 번역된 전체 텍스트
full_text = "\n".join(s["text"] for s in translated_segments)
self.update_state(state="PROGRESS", meta={"progress":93,"message":"파일 저장 중..."})
os.makedirs(OUTPUT_DIR, exist_ok=True)
result_files = {}
# ── TXT 저장 ─────────────────────────────────────────
txt_filename = f"{file_id}.txt"
with open(os.path.join(OUTPUT_DIR, txt_filename), "w", encoding="utf-8") as f:
f.write(f"# 변환 결과\n")
f.write(f"# 언어: {detected_lang} | 재생 시간: {duration:.1f}\n")
if is_translated:
f.write(f"# 번역: {_lang_name(translate_to)}\n")
f.write(f"\n## 전체 텍스트\n\n{full_text}\n\n")
f.write(f"## 타임스탬프별 세그먼트\n\n")
for seg in (translated_segments if is_translated else segments):
f.write(f"[{_fmt_ts(seg['start'])}{_fmt_ts(seg['end'])}] {seg['text']}\n")
result_files["txt"] = txt_filename
# ── 자막 파일 저장 ────────────────────────────────────
if subtitle_mode:
sub_segs = translated_segments if is_translated else segments
lang_suffix = f".{translate_to}" if is_translated else f".{detected_lang}"
if subtitle_format in ("srt", "both"):
srt_fn = f"{file_id}{lang_suffix}.srt"
with open(os.path.join(OUTPUT_DIR, srt_fn), "w", encoding="utf-8") as f:
f.write(_make_srt(sub_segs))
result_files["srt"] = srt_fn
if subtitle_format in ("vtt", "both"):
vtt_fn = f"{file_id}{lang_suffix}.vtt"
with open(os.path.join(OUTPUT_DIR, vtt_fn), "w", encoding="utf-8") as f:
f.write(_make_vtt(sub_segs))
result_files["vtt"] = vtt_fn
# 원본 언어 SRT도 함께 (번역 시)
if is_translated and subtitle_format in ("srt","both"):
orig_fn = f"{file_id}.{detected_lang}.srt"
with open(os.path.join(OUTPUT_DIR, orig_fn), "w", encoding="utf-8") as f:
f.write(_make_srt(segments))
result_files["srt_original"] = orig_fn
try: os.remove(audio_path)
except: pass
return {
# 기본 STT 결과
"text": full_text,
"raw_text": raw_text,
"segments": translated_segments if is_translated else segments,
"orig_segments": segments,
"language": detected_lang,
"duration": round(duration, 1),
# 후처리
"ollama_used": use_ollama and bool(ollama_model) and not subtitle_mode,
"ollama_model": ollama_model if (use_ollama and not subtitle_mode) else "",
"openrouter_used": use_openrouter and bool(openrouter_model) and not subtitle_mode,
"openrouter_model": openrouter_model if (use_openrouter and not subtitle_mode) else "",
# 자막
"subtitle_mode": subtitle_mode,
"subtitle_format": subtitle_format,
"translated": is_translated,
"translate_to": translate_to if is_translated else "",
"translate_model": translate_model if is_translated else "",
# 파일
"output_file": result_files.get("txt",""),
"srt_file": result_files.get("srt",""),
"vtt_file": result_files.get("vtt",""),
"srt_original_file": result_files.get("srt_original",""),
}
except Exception as e:
raise Exception(f"변환 실패: {str(e)}")
def _fmt_ts(s: float) -> str:
m, sec = divmod(int(s), 60)
return f"{m:02d}:{sec:02d}"