File size: 12,941 Bytes
a0fa1cf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 | """
Forced alignment for the essay-to-video pipeline (Batch 2).
Given the essay's existing narration audio and the script-mode scenes (whose
narration is the essay's own sentences, verbatim - Batch 0 guarantees this),
compute where each scene's words actually fall in the audio. Two things fall
out: real per-scene durations (cutting on sentence boundaries, not the flat
5-second default) and word-level timestamps persisted for caption
materialization in the promote-to-project bridge (Batch 4) - no re-run needed.
Two alignment tiers, best available wins:
whisperx - word-level timestamps via WhisperX (Whisper + wav2vec2).
Heavy optional dependency; imported lazily and only when
installed. Env: WHISPERX_MODEL (default large-v3),
WHISPERX_DEVICE (default cuda; falls back to cpu).
proportional - deterministic fallback: total audio duration (ffprobe, or
the wave module for WAV) distributed across scenes by word
count. No dependencies. Scene boundaries are already
sentence boundaries, so cuts stay clean - just less exact.
Alignment is strictly opt-in: no audio URL, no alignment, and the Batch 0
flat-duration behavior is untouched.
"""
from __future__ import annotations
import json
import os
import re
import shutil
import subprocess
import tempfile
import wave
from pathlib import Path
from typing import List, Optional
from pydantic import BaseModel, Field
class AlignedWord(BaseModel):
word: str
start: float
end: float
class SceneSpan(BaseModel):
scene_number: int
start_sec: float
end_sec: float
@property
def duration_sec(self) -> float:
return max(0.0, self.end_sec - self.start_sec)
class AlignmentResult(BaseModel):
ok: bool
method: str = "" # "whisperx" | "proportional"
audio_duration_sec: float = 0.0
words: List[AlignedWord] = Field(default_factory=list)
scene_spans: List[SceneSpan] = Field(default_factory=list)
message: str = ""
# ββ Audio helpers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def fetch_audio(url_or_path: str, timeout: float = 60.0) -> str:
"""Return a local file path for the audio, downloading if it's a URL.
Local filesystem paths are refused unless STUDIO_ALLOW_LOCAL_AUDIO=true:
existing_audio_url is API-supplied, and quietly reading server paths from
it would be an SSRF-style hole in multi-user deployments. Single-user
installs that want to point at local files opt in explicitly."""
if not url_or_path.startswith(("http://", "https://")):
allow_local = os.getenv("STUDIO_ALLOW_LOCAL_AUDIO", "false").strip().lower() in (
"1", "true", "yes")
if not allow_local:
raise PermissionError(
"Local audio paths are disabled (set STUDIO_ALLOW_LOCAL_AUDIO=true "
"on trusted single-user installs); pass an http(s) URL instead.")
if not Path(url_or_path).is_file():
raise FileNotFoundError(url_or_path)
return url_or_path
import httpx
suffix = Path(url_or_path.split("?", 1)[0]).suffix or ".mp3"
fd, path = tempfile.mkstemp(suffix=suffix, prefix="essay-audio-")
os.close(fd)
with httpx.stream("GET", url_or_path, timeout=timeout, follow_redirects=True) as resp:
resp.raise_for_status()
with open(path, "wb") as fh:
for chunk in resp.iter_bytes():
fh.write(chunk)
return path
def audio_duration_sec(path: str) -> Optional[float]:
"""Duration via ffprobe, falling back to the wave module for WAV files."""
ffprobe = shutil.which("ffprobe")
if ffprobe:
try:
out = subprocess.run(
[ffprobe, "-v", "quiet", "-print_format", "json",
"-show_format", path],
capture_output=True, check=True, timeout=30,
)
dur = float(json.loads(out.stdout).get("format", {}).get("duration", 0))
if dur > 0:
return dur
except Exception:
pass
if path.lower().endswith(".wav"):
try:
with wave.open(path, "rb") as wf:
return wf.getnframes() / float(wf.getframerate())
except Exception:
pass
return None
# ββ WhisperX tier ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def whisperx_available() -> bool:
try:
import whisperx # noqa: F401
return True
except Exception:
return False
def _whisperx_words(audio_path: str) -> List[AlignedWord]:
"""Transcribe + align, returning word-level timestamps."""
import whisperx
device = os.getenv("WHISPERX_DEVICE", "cuda").strip() or "cuda"
model_name = os.getenv("WHISPERX_MODEL", "large-v3").strip() or "large-v3"
try:
model = whisperx.load_model(model_name, device)
except Exception:
device = "cpu"
model = whisperx.load_model(model_name, device, compute_type="int8")
audio = whisperx.load_audio(audio_path)
result = model.transcribe(audio)
align_model, align_meta = whisperx.load_align_model(
language_code=result["language"], device=device)
aligned = whisperx.align(result["segments"], align_model, align_meta,
audio, device)
words: List[AlignedWord] = []
for seg in aligned.get("word_segments", []):
w = str(seg.get("word", "")).strip()
if w and seg.get("start") is not None and seg.get("end") is not None:
words.append(AlignedWord(word=w, start=float(seg["start"]),
end=float(seg["end"])))
return words
# ββ Scene span mapping βββββββββββββββββββββββββββββββββββββββββββββββββββββββ
_WORD_RE = re.compile(r"[\w']+")
def _word_count(text: str) -> int:
return len(_WORD_RE.findall(text))
def _spans_from_words(scenes: List[dict], words: List[AlignedWord],
audio_duration: float) -> List[SceneSpan]:
"""
Map scenes onto the aligned word list by cumulative word position.
The narration is the same text the audio narrates (Batch 0's verbatim
guarantee), but the transcript may differ slightly (numbers, hyphens),
so exact token matching is brittle. Cumulative-proportional indexing
into the *aligned* word list keeps boundaries on real word timestamps
while tolerating small transcription drift.
"""
counts = [max(1, _word_count(s.get("narration", ""))) for s in scenes]
total = sum(counts)
n_words = len(words)
spans: List[SceneSpan] = []
cursor = 0
consumed = 0
for i, scene in enumerate(scenes):
consumed += counts[i]
end_idx = min(n_words - 1, round(consumed / total * n_words) - 1)
start_sec = words[cursor].start if cursor < n_words else audio_duration
end_sec = words[end_idx].end if end_idx >= cursor else start_sec
if i == len(scenes) - 1:
end_sec = max(end_sec, audio_duration)
spans.append(SceneSpan(scene_number=scene.get("scene_number", i + 1),
start_sec=round(start_sec, 3),
end_sec=round(end_sec, 3)))
cursor = min(n_words - 1, end_idx + 1)
return spans
def _spans_proportional(scenes: List[dict], audio_duration: float) -> List[SceneSpan]:
counts = [max(1, _word_count(s.get("narration", ""))) for s in scenes]
total = sum(counts)
spans: List[SceneSpan] = []
t = 0.0
for i, scene in enumerate(scenes):
dur = audio_duration * counts[i] / total
end = audio_duration if i == len(scenes) - 1 else t + dur
spans.append(SceneSpan(scene_number=scene.get("scene_number", i + 1),
start_sec=round(t, 3), end_sec=round(end, 3)))
t = end
return spans
# ββ Entry points βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def align_scenes(audio_path: str, scenes: List[dict]) -> AlignmentResult:
"""Align scenes against a local audio file. Never raises for missing
optional dependencies - degrades to the proportional tier."""
duration = audio_duration_sec(audio_path)
if not duration or duration <= 0:
return AlignmentResult(ok=False, message="Could not determine audio duration "
"(is ffprobe installed?)")
if whisperx_available():
try:
words = _whisperx_words(audio_path)
if words:
return AlignmentResult(
ok=True, method="whisperx", audio_duration_sec=duration,
words=words, scene_spans=_spans_from_words(scenes, words, duration),
)
except Exception as e:
print(f"[Alignment] whisperx failed ({e}); falling back to proportional")
return AlignmentResult(
ok=True, method="proportional", audio_duration_sec=duration,
scene_spans=_spans_proportional(scenes, duration),
)
def align_from_url(audio_url: str, scenes: List[dict]) -> AlignmentResult:
"""Fetch audio (if remote) and align. Cleans up its own temp download."""
path = fetch_audio(audio_url)
is_temp = path != audio_url
try:
return align_scenes(path, scenes)
finally:
if is_temp:
try:
os.unlink(path)
except OSError:
pass
class CaptionCue(BaseModel):
start_sec: float
end_sec: float
text: str
def caption_cues(result: AlignmentResult, scenes: List[dict],
max_chars: int = 42) -> List[CaptionCue]:
"""
Build caption cues from a persisted AlignmentResult - the data source
for CaptionSegment rows in promote-to-project (Batch 4), with no
re-run of alignment.
whisperx tier: group aligned words into <= max_chars cues on real word
timestamps. proportional tier: chunk each scene's narration and spread
the chunks evenly across the scene's span.
"""
cues: List[CaptionCue] = []
if result.words:
line: List[str] = []
start = result.words[0].start
for w in result.words:
probe = " ".join(line + [w.word])
if line and len(probe) > max_chars:
cues.append(CaptionCue(start_sec=round(start, 3),
end_sec=round(w.start, 3),
text=" ".join(line)))
line, start = [w.word], w.start
else:
line.append(w.word)
if line:
cues.append(CaptionCue(start_sec=round(start, 3),
end_sec=round(result.words[-1].end, 3),
text=" ".join(line)))
return cues
by_number = {sp.scene_number: sp for sp in result.scene_spans}
for i, scene in enumerate(scenes):
span = by_number.get(scene.get("scene_number", i + 1))
text = (scene.get("narration") or "").strip()
if not span or not text:
continue
words = text.split()
chunks: List[str] = []
line: List[str] = []
for w in words:
if line and len(" ".join(line + [w])) > max_chars:
chunks.append(" ".join(line))
line = [w]
else:
line.append(w)
if line:
chunks.append(" ".join(line))
step = span.duration_sec / max(1, len(chunks))
for j, chunk in enumerate(chunks):
cues.append(CaptionCue(
start_sec=round(span.start_sec + j * step, 3),
end_sec=round(span.start_sec + (j + 1) * step, 3),
text=chunk))
return cues
def apply_spans_to_scenes(scenes: List[dict], result: AlignmentResult) -> int:
"""Write real durations (and audio offsets, for Batch 4 caption/audio
slicing) into outline scene dicts. Returns how many scenes were updated."""
by_number = {sp.scene_number: sp for sp in result.scene_spans}
updated = 0
for i, scene in enumerate(scenes):
span = by_number.get(scene.get("scene_number", i + 1))
if not span or span.duration_sec <= 0:
continue
scene["duration_sec"] = round(span.duration_sec, 3)
scene["audio_start_sec"] = span.start_sec
scene["audio_end_sec"] = span.end_sec
updated += 1
return updated
|