# ───────────────────────────────────────────── # NINA — embedded subtitle extraction # ───────────────────────────────────────────── # When the audio is too garbled for reliable ASR, a clean human-authored subtitle # track (if the upload carries one) is far better than anything we can transcribe. # This reads text-based subtitle streams straight out of the container with ffmpeg # — no new dependencies. Bitmap subtitles (PGS/VobSub) and burned-in captions are # NOT handled here; those need OCR (a separate, heavier step). # # Best-effort: returns [] on anything unexpected, so the pipeline simply falls # back to the ASR transcript. import json import os import re import shutil import subprocess import tempfile from config import SUBTITLE_TEXT_CODECS _SRT_TIME = re.compile( r"(\d+):(\d+):(\d+)[,.](\d+)\s*-->\s*(\d+):(\d+):(\d+)[,.](\d+)" ) def _subtitle_stream_index(video_path: str): """Return the 0-based index (among subtitle streams) of the first text-based subtitle track, or None.""" try: out = subprocess.run( ["ffprobe", "-v", "error", "-select_streams", "s", "-show_entries", "stream=index,codec_name", "-of", "json", video_path], capture_output=True, text=True, timeout=30, ) streams = (json.loads(out.stdout or "{}")).get("streams", []) for i, s in enumerate(streams): if s.get("codec_name", "").lower() in SUBTITLE_TEXT_CODECS: return i except Exception: pass return None def _parse_srt(text: str) -> list[dict]: segments: list[dict] = [] for block in re.split(r"\n\s*\n", text.strip()): m = _SRT_TIME.search(block) if not m: continue h1, m1, s1, ms1, h2, m2, s2, ms2 = map(int, m.groups()) start = h1 * 3600 + m1 * 60 + s1 + ms1 / 1000.0 end = h2 * 3600 + m2 * 60 + s2 + ms2 / 1000.0 lines = block.splitlines() ti = next((i for i, ln in enumerate(lines) if "-->" in ln), 0) txt = " ".join(ln.strip() for ln in lines[ti + 1:] if ln.strip()) txt = re.sub(r"<[^>]+>", "", txt) # strip styling tags txt = re.sub(r"\{[^}]+\}", "", txt) # strip ASS override blocks txt = txt.strip() if txt: segments.append({"text": txt, "start": round(start, 2), "end": round(end, 2)}) return segments def extract_embedded_subs(video_path: str) -> list[dict]: """Extract the first text-based subtitle track as [{text,start,end}], or [].""" try: if not (shutil.which("ffmpeg") and shutil.which("ffprobe")): return [] idx = _subtitle_stream_index(video_path) if idx is None: return [] fd, srt = tempfile.mkstemp(prefix="nina_subs_", suffix=".srt") os.close(fd) try: r = subprocess.run( ["ffmpeg", "-y", "-i", video_path, "-map", f"0:s:{idx}", "-f", "srt", srt, "-loglevel", "error"], capture_output=True, timeout=120, ) if r.returncode != 0: return [] with open(srt, encoding="utf-8", errors="ignore") as f: segs = _parse_srt(f.read()) if segs: print(f"[NINA] Embedded subtitle track found — {len(segs)} cues.") return segs finally: try: os.remove(srt) except OSError: pass except Exception as e: print(f"[NINA] Subtitle extraction failed: {e}") return []