39ba5ddf76
Schema changes: dev.chunks->dev.chunk, remove old_chunk_id/chunk_index Correction: asr-1.json format, generate/apply scripts API: 37/37 endpoints fixed and tested Docs: HANDOVER_V2.0.md for M4
99 lines
3.2 KiB
Python
99 lines
3.2 KiB
Python
#!/opt/homebrew/bin/python3.11
|
|
"""
|
|
Step 3: Re-run ASR with word_timestamps on full audio.
|
|
Map words to 4188 fine segments for accurate text.
|
|
"""
|
|
import json, sys, os, time, subprocess, tempfile, shutil
|
|
from faster_whisper import WhisperModel
|
|
|
|
UUID = "aeed71342a899fe4b4c57b7d41bcb692"
|
|
BASE = "/Users/accusys/momentry/output_dev"
|
|
VIDEO = "/Users/accusys/momentry/var/sftpgo/data/demo/Charade (1963) Cary Grant & Audrey Hepburn \uff5c Comedy Mystery Romance Thriller \uff5c Full Movie.mp4"
|
|
|
|
print("=== Load fine ASRX ===")
|
|
fine = json.load(open(f"{BASE}/{UUID}.asrx_fine.json"))
|
|
fine_segs = fine["segments"]
|
|
print(f"Fine segments: {len(fine_segs)}")
|
|
|
|
print("\n=== Extract audio WAV ===")
|
|
tmp_dir = tempfile.mkdtemp(prefix="asr_step3_")
|
|
wav_path = os.path.join(tmp_dir, "audio.wav")
|
|
subprocess.run(["ffmpeg", "-y", "-v", "quiet", "-i", VIDEO,
|
|
"-ar", "16000", "-ac", "1", "-sample_fmt", "s16", wav_path],
|
|
check=True, capture_output=True, timeout=300)
|
|
|
|
print("Loading model with word_timestamps...")
|
|
t0 = time.time()
|
|
model = WhisperModel("small", device="cpu", compute_type="int8")
|
|
print(f" Model loaded in {time.time()-t0:.1f}s")
|
|
|
|
print("Transcribing with word_timestamps=True...")
|
|
t0 = time.time()
|
|
segments, info = model.transcribe(
|
|
wav_path, beam_size=5, vad_filter=True,
|
|
vad_parameters={"min_silence_duration_ms": 500},
|
|
word_timestamps=True
|
|
)
|
|
|
|
# Collect all word-level data
|
|
words = []
|
|
for seg in segments:
|
|
if seg.words:
|
|
for w in seg.words:
|
|
wt = w.word.strip()
|
|
if wt:
|
|
words.append({"word": wt, "start": w.start, "end": w.end})
|
|
else:
|
|
words.append({"word": seg.text.strip(), "start": seg.start, "end": seg.end})
|
|
|
|
elapsed = time.time() - t0
|
|
print(f" Done in {elapsed:.1f}s, {len(words)} words")
|
|
|
|
# Map words to fine segments
|
|
print("\n=== Map words to fine segments ===")
|
|
wi = 0
|
|
assigned = 0
|
|
for si, fs in enumerate(fine_segs):
|
|
fstart = fs["start_time"]
|
|
fend = fs["end_time"]
|
|
seg_words = []
|
|
|
|
while wi < len(words):
|
|
w = words[wi]
|
|
if w["end"] <= fstart:
|
|
wi += 1
|
|
continue
|
|
if w["start"] >= fend:
|
|
break
|
|
seg_words.append(w["word"])
|
|
wi += 1
|
|
|
|
text = " ".join(seg_words)
|
|
fs["text"] = text
|
|
if text:
|
|
assigned += 1
|
|
|
|
print(f" Segments with text: {assigned}/{len(fine_segs)}")
|
|
|
|
# Show examples
|
|
print("\nSplit segment examples:")
|
|
for fs in fine_segs:
|
|
# Check if this was split (doesn't match an ASR boundary exactly)
|
|
is_split = True
|
|
# We can't easily check here, just show first 10 non-trivial
|
|
if len(fs.get('text','')) > 10 and is_split:
|
|
print(f" [{fs['start_time']:.1f}-{fs['end_time']:.1f}] {fs['speaker_name']:15s} \"{fs['text'][:60]}\"")
|
|
break # just one for now
|
|
|
|
# Count text lengths
|
|
text_lens = [len(fs.get('text','')) for fs in fine_segs]
|
|
print(f"\n Avg text length: {sum(text_lens)/len(text_lens):.0f} chars")
|
|
print(f" Empty texts: {sum(1 for l in text_lens if l == 0)}")
|
|
|
|
# Save
|
|
fine["_asr_meta"] = {"word_timestamps": True, "asr_runtime_secs": round(elapsed, 1)}
|
|
json.dump(fine, open(f"{BASE}/{UUID}.asrx_fine.json", "w"), indent=2)
|
|
print(f"\nSaved")
|
|
|
|
shutil.rmtree(tmp_dir, ignore_errors=True)
|