#!/usr/bin/env python3
"""Word timestamps for takes, from a local Whisper — the cutdown cuts INSIDE lines, and the QA
listener's timestamps drift too much for that.

    uv run --with faster-whisper python tools/trailer/words.py <take|file>...  (take names in promo/trailer/clips)
"""
import os, subprocess, sys
import numpy as np
from faster_whisper import WhisperModel

ROOT = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
m = WhisperModel('small.en', compute_type='int8')
for name in sys.argv[1:]:
    # decode with ffmpeg ourselves: faster-whisper's PyAV path breaks on newer av releases
    src = name if os.path.exists(name) else os.path.join(ROOT, 'promo', 'trailer', 'clips', name + '.mp4')
    raw = subprocess.run(['ffmpeg', '-v', 'error', '-i', src,
                          '-ac', '1', '-ar', '16000', '-f', 's16le', '-'], capture_output=True).stdout
    segs, _ = m.transcribe(np.frombuffer(raw, np.int16).astype(np.float32) / 32768, word_timestamps=True)
    print(name + ':', '  '.join(f'{w.word.strip()}[{w.start:.2f}-{w.end:.2f}]' for s in segs for w in s.words))