feat(web): speaker diarization via pyannote (parity with whisper_stt)

- Add app/diarize.py: local snapshot, A/B labels, disclaimer text
- transcribe_file and async jobs support diarize flag; Form diarize on API
- UI checkbox (default on); requirements: pyannote.audio, huggingface_hub
- README: env vars and model notes

Made-with: Cursor

app/diarize.py

@@ -0,0 +1,172 @@
+"""웹 STT용 화자 분리 — whisper_stt.py와 동일한 pyannote 로컬 스냅샷 + 타임라인 정렬."""
+from __future__ import annotations
+
+import os
+from pathlib import Path
+from typing import Any
+
+from .pyannote_auth import load_pyannote_pipeline
+
+PROJECT_ROOT = Path(__file__).resolve().parent.parent
+DEFAULT_DIARIZE_MODEL_DIR = PROJECT_ROOT / "models" / "pyannote-diarization-3.1"
+
+DIARIZE_DISCLAIMER_KO = (
+    "※ 화자 A, B, C… 는 실제 이름이 아니라, 이 녹음에서 말이 처음 잡힌 순서로 붙인 구분자입니다.\n"
+    "※ 같은 사람이 여러 구간으로 나뉘면 라벨이 바뀌거나 섞일 수 있으니, 중요한 회의는 검수가 필요합니다.\n\n"
+)
+
+
+def diarization_annotation(diarization: Any) -> Any:
+    """pyannote.audio 4.x는 DiarizeOutput; 구간은 .speaker_diarization에 있다."""
+    ann = getattr(diarization, "speaker_diarization", None)
+    return diarization if ann is None else ann
+
+
+def validate_pyannote_snapshot(model_dir: Path | str) -> None:
+    cfg = Path(model_dir) / "config.yaml"
+    if cfg.is_file():
+        return
+    p = Path(model_dir).resolve()
+    raise ValueError(
+        f"pyannote 모델 폴더가 불완전합니다 (config.yaml 없음): {p}. "
+        "hf download pyannote/speaker-diarization-3.1 --local-dir ./models/pyannote-diarization-3.1"
+    )
+
+
+def resolve_local_diarize_dir(override: str | None) -> Path:
+    if override:
+        path = Path(override).expanduser().resolve()
+        if path.is_dir():
+            return path
+        raise ValueError(f"화자 분리 모델 폴더가 없습니다: {path}")
+
+    for cand in (os.environ.get("WHISPER_DIARIZE_MODEL_DIR"), os.environ.get("PYANNOTE_MODEL_DIR")):
+        if cand:
+            path = Path(cand).expanduser().resolve()
+            if path.is_dir():
+                return path
+
+    path = DEFAULT_DIARIZE_MODEL_DIR.resolve()
+    if path.is_dir():
+        return path
+    raise ValueError(
+        f"화자 분리 모델 폴더가 없습니다: {path}. "
+        "프로젝트 루트에서: hf download pyannote/speaker-diarization-3.1 "
+        "--local-dir ./models/pyannote-diarization-3.1 (약관 동의·HF 토큰 필요)"
+    )
+
+
+def speaker_turns(audio_path: str, *, model_dir: str | None = None) -> list[tuple[float, float, str]]:
+    import torch
+
+    resolved = resolve_local_diarize_dir(model_dir)
+    validate_pyannote_snapshot(resolved)
+    pipeline = load_pyannote_pipeline(resolved)
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    pipeline.to(device)
+    diarization = pipeline(audio_path)
+    ann = diarization_annotation(diarization)
+    turns: list[tuple[float, float, str]] = []
+    for segment, _, label in ann.itertracks(yield_label=True):
+        turns.append((float(segment.start), float(segment.end), str(label)))
+    turns.sort(key=lambda x: x[0])
+    return turns
+
+
+def _overlap_sec(a0: float, a1: float, b0: float, b1: float) -> float:
+    return max(0.0, min(a1, b1) - max(a0, b0))
+
+
+def _assign_speaker(
+    seg_start: float, seg_end: float, turns: list[tuple[float, float, str]]
+) -> str | None:
+    best: str | None = None
+    best_ov = 0.0
+    for t0, t1, sp in turns:
+        ov = _overlap_sec(seg_start, seg_end, t0, t1)
+        if ov > best_ov:
+            best_ov = ov
+            best = sp
+    if best is None or best_ov < 0.05:
+        return None
+    return best
+
+
+def speaker_label_order(turns: list[tuple[float, float, str]]) -> dict[str, str]:
+    order: list[str] = []
+    for t0, _, sp in sorted(turns, key=lambda x: x[0]):
+        if sp not in order:
+            order.append(sp)
+
+    def letter(i: int) -> str:
+        if i < 26:
+            return chr(ord("A") + i)
+        return f"SP{i + 1}"
+
+    return {sp: letter(i) for i, sp in enumerate(order)}
+
+
+def format_diarized_text(
+    whisper_segments: list[dict[str, Any]],
+    turns: list[tuple[float, float, str]],
+) -> str:
+    labels = speaker_label_order(turns)
+    lines: list[str] = []
+    current_letter: str | None = None
+    current_parts: list[str] = []
+
+    def flush() -> None:
+        nonlocal current_letter, current_parts
+        if current_letter is not None and current_parts:
+            lines.append(f"{current_letter}: {' '.join(current_parts).strip()}")
+        current_letter = None
+        current_parts = []
+
+    for seg in whisper_segments:
+        text = (seg.get("text") or "").strip()
+        if not text:
+            continue
+        start = float(seg["start"])
+        end = float(seg["end"])
+        sp = _assign_speaker(start, end, turns)
+        letter = labels.get(sp, "?") if sp is not None else "?"
+
+        if letter == current_letter:
+            current_parts.append(text)
+        else:
+            flush()
+            current_letter = letter
+            current_parts = [text]
+
+    flush()
+    return "\n".join(lines)
+
+
+def segments_with_speakers(
+    whisper_segments: list[dict[str, Any]],
+    turns: list[tuple[float, float, str]],
+) -> list[dict[str, Any]]:
+    labels = speaker_label_order(turns)
+    out: list[dict[str, Any]] = []
+    for seg in whisper_segments:
+        text = (seg.get("text") or "").strip()
+        if not text:
+            continue
+        sp = _assign_speaker(float(seg["start"]), float(seg["end"]), turns)
+        letter = labels.get(sp, "?") if sp is not None else "?"
+        out.append({**seg, "text": text, "speaker": letter})
+    return out
+
+
+def build_diarized_output(
+    whisper_segments: list[dict[str, Any]],
+    audio_path: str,
+    *,
+    model_dir: str | None = None,
+    with_disclaimer: bool = True,
+) -> tuple[str, list[dict[str, Any]]]:
+    turns = speaker_turns(audio_path, model_dir=model_dir)
+    body = format_diarized_text(whisper_segments, turns)
+    text = (DIARIZE_DISCLAIMER_KO + body) if with_disclaimer else body
+    segs = segments_with_speakers(whisper_segments, turns)
+    return text, segs

app/main.py

@@ -61,6 +61,7 @@ class _Job:
     language: str | None
     vad_filter: bool
     beam_size: int
+    diarize: bool
     author_id: str
     language_requested: str | None
     status: str = "queued"  # queued|running|completed|failed|cancelled
@@ -128,6 +129,7 @@ async def api_create_job(
     language: str = Form(default="ko"),
     vad_filter: bool = Form(default=True),
     beam_size: int = Form(default=5),
+    diarize: bool = Form(default=True),
     author_id: str = Form(default=DEFAULT_AUTHOR_ID),
 ) -> dict[str, Any]:
     _cleanup_jobs()
@@ -146,6 +148,7 @@ async def api_create_job(
         language=(lang or None),
         vad_filter=bool(vad_filter),
         beam_size=int(beam_size),
+        diarize=bool(diarize),
         author_id=(author_id.strip() or DEFAULT_AUTHOR_ID),
         language_requested=(language.strip() or None),
         status="queued",
@@ -188,6 +191,7 @@ async def api_transcribe(
     language: str = Form(default="ko"),
     vad_filter: bool = Form(default=True),
     beam_size: int = Form(default=5),
+    diarize: bool = Form(default=True),
     author_id: str = Form(default=DEFAULT_AUTHOR_ID),
 ) -> dict[str, Any]:
     _validate_upload(file)
@@ -203,6 +207,7 @@ async def api_transcribe(
             language=(lang or None),
             vad_filter=bool(vad_filter),
             beam_size=int(beam_size),
+            diarize=bool(diarize),
         )
         # 단발성 API도 DB 저장
         try:
@@ -357,6 +362,7 @@ def _run_job(job_id: str) -> None:
             language = job.language
             vad_filter = job.vad_filter
             beam_size = job.beam_size
+            do_diarize = job.diarize
             author_id = job.author_id
             language_requested = job.language_requested
             filename = job.filename
@@ -421,6 +427,34 @@ def _run_job(job_id: str) -> None:
                     job.progress = None
                 job.updated_at = time.time()
 
+        if not cancelled and do_diarize:
+            with _JOBS_LOCK:
+                job = _JOBS.get(job_id)
+                if job is None:
+                    return
+                if job.cancel_event.is_set():
+                    cancelled = True
+                else:
+                    segs_snapshot = [dict(s) for s in job.segments]
+                    path_for_diar = job.tmp_path
+
+            if not cancelled and segs_snapshot:
+                from . import diarize as dz
+
+                mdir = os.getenv("APP_PYANNOTE_MODEL_DIR") or None
+                with _JOBS_LOCK:
+                    job = _JOBS.get(job_id)
+                    if job is not None:
+                        job.progress = 0.97
+                        job.updated_at = time.time()
+                text_d, segs_d = dz.build_diarized_output(segs_snapshot, path_for_diar, model_dir=mdir)
+                with _JOBS_LOCK:
+                    job = _JOBS.get(job_id)
+                    if job is not None:
+                        job.text = text_d
+                        job.segments = segs_d
+                        job.updated_at = time.time()
+
         with _JOBS_LOCK:
             job = _JOBS.get(job_id)
             if job is None:

app/static/index.html

@@ -305,6 +305,11 @@
             VAD 필터 (무음 구간 감소)
           </label>
 
+          <label>
+            <input id="diarize" type="checkbox" checked />
+            화자 분리 (pyannote, whisper_stt.py와 동일 방식 — 서버에 로컬 모델·HF 토큰 필요)
+          </label>
+
           <div class="row" style="margin-top: 12px">
             <button class="btn primary" id="go" disabled>전사(STT) 실행</button>
             <button class="btn" id="cancel" disabled>취소</button>
@@ -314,7 +319,8 @@
 
           <div class="hint">
             - 허용: mp3, m4a, wav, mp4, aac, ogg, flac, webm<br />
-            - 첫 실행 시 Whisper 모델 다운로드로 시간이 걸릴 수 있습니다.
+            - 첫 실행 시 Whisper 모델 다운로드로 시간이 걸릴 수 있습니다.<br />
+            - 화자 분리 켜짐: <span class="mono">./models/pyannote-diarization-3.1</span> 및 gated HF 모델 동의(README 참고).
           </div>
 
           <div class="progress">
@@ -420,6 +426,7 @@
       const progTextEl = $("progText");
       const downloadEl = $("download");
       const clearEl = $("clear");
+      const diarizeEl = $("diarize");
       const healthEl = $("health");
       const metaEl = $("meta");
       const timingEl = $("timing");
@@ -677,6 +684,7 @@
           const author = (authorEl?.value || "").trim();
           if (author) fd.append("author_id", author);
           fd.append("vad_filter", $("vad").checked ? "true" : "false");
+          fd.append("diarize", !diarizeEl || diarizeEl.checked ? "true" : "false");
           fd.append("beam_size", $("beam").value);
 
           uploadController = new AbortController();

app/stt.py

@@ -54,6 +54,8 @@ def transcribe_file(
     language: str | None = None,
     vad_filter: bool = True,
     beam_size: int = 5,
+    diarize: bool = True,
+    diarize_model_dir: str | None = None,
 ) -> dict[str, Any]:
     segments_iter, info = transcribe_iter(
         audio_path,
@@ -70,10 +72,23 @@ def transcribe_file(
             segments.append(seg)
             texts.append(seg.text)
 
+    seg_dicts = [seg.__dict__ for seg in segments]
     full_text = "\n".join(texts).strip()
+
+    if diarize:
+        from . import diarize as dz
+
+        mdir = diarize_model_dir or os.getenv("APP_PYANNOTE_MODEL_DIR") or None
+        full_text, seg_dicts = dz.build_diarized_output(
+            seg_dicts,
+            audio_path,
+            model_dir=mdir,
+            with_disclaimer=True,
+        )
+
     return {
         "text": full_text,
-        "segments": [seg.__dict__ for seg in segments],
+        "segments": seg_dicts,
         "detected_language": getattr(info, "language", None),
         "language_probability": getattr(info, "language_probability", None),
         "duration_sec": getattr(info, "duration", None),