91 lines
4.1 KiB
Python
91 lines
4.1 KiB
Python
"""Split a voice batch against word timestamps; reject missing script words.
|
||
|
||
Usage: python split_voice_batch.py AUDIO WORDS_JSON LINES_JSON OUTPUT_DIRECTORY
|
||
Requires soundfile (the local build helper directory is also supported).
|
||
"""
|
||
from pathlib import Path
|
||
import difflib
|
||
import json
|
||
import re
|
||
import sys
|
||
|
||
ROOT = Path(__file__).resolve().parents[1]
|
||
sys.path.insert(0, str(ROOT / "builds/python_audio"))
|
||
import soundfile as sf
|
||
|
||
DIGITS = dict(zip("0123456789", "ноль один два три четыре пять шесть семь восемь девять".split()))
|
||
|
||
def tokens(text):
|
||
result = []
|
||
for token in re.findall(r"[а-яёa-z0-9]+", text.lower().replace("ё", "е")):
|
||
result.extend([DIGITS[d] for d in token] if token.isdigit() else [token])
|
||
return result
|
||
|
||
def main():
|
||
audio_path, words_path, lines_path, output_path = map(Path, sys.argv[1:5])
|
||
data = json.loads(words_path.read_text(encoding="utf-8"))
|
||
words = data["words"] if isinstance(data, dict) else data
|
||
lines = json.loads(lines_path.read_text(encoding="utf-8"))
|
||
spoken, timing = [], []
|
||
for word in words:
|
||
if word.get("type", "word") != "word":
|
||
continue
|
||
for token in tokens(word["text"]):
|
||
spoken.append(token)
|
||
timing.append((float(word["start"]), float(word["end"])))
|
||
reference, spans = [], []
|
||
for line in lines:
|
||
start = len(reference)
|
||
reference.extend(tokens(line["text"]))
|
||
spans.append((start, len(reference)))
|
||
aligned = {}
|
||
matcher = difflib.SequenceMatcher(None, reference, spoken, autojunk=False)
|
||
for a, b, length in matcher.get_matching_blocks():
|
||
for i in range(length):
|
||
aligned[a + i] = b + i
|
||
mismatches = []
|
||
for tag, a, end_a, b, end_b in matcher.get_opcodes():
|
||
if tag != "equal":
|
||
mismatches.append(dict(kind=tag, reference=" ".join(reference[a:end_a]),
|
||
spoken=" ".join(spoken[b:end_b]), at_reference=a))
|
||
print(json.dumps({"reference_words": len(reference), "spoken_words": len(spoken),
|
||
"matched": len(aligned), "mismatches": mismatches}, ensure_ascii=False))
|
||
missing = [(line["id"], " ".join(reference[a:b])) for line, (a, b) in zip(lines, spans)
|
||
if any(i not in aligned for i in range(a, b))]
|
||
if missing:
|
||
raise SystemExit("Missing script words; no clips written: " + repr(missing))
|
||
boundaries = {point for span in spans for point in span}
|
||
if any(m["kind"] == "insert" and m["at_reference"] not in boundaries for m in mismatches):
|
||
raise SystemExit("Unexpected words inside a line; review the transcript before exporting")
|
||
samples, rate = sf.read(audio_path, always_2d=True)
|
||
samples = samples.mean(axis=1)
|
||
output_path.mkdir(parents=True, exist_ok=True)
|
||
report = []
|
||
for line, (a, b) in zip(lines, spans):
|
||
first, last = aligned[a], aligned[b - 1]
|
||
start = max(0, timing[first][0] - .10)
|
||
end = min(len(samples) / rate, timing[last][1] + .16)
|
||
if first > 0:
|
||
start = max(start, (timing[first - 1][1] + timing[first][0]) / 2)
|
||
if last + 1 < len(timing):
|
||
end = min(end, (timing[last][1] + timing[last + 1][0]) / 2)
|
||
clip = samples[round(start * rate):round(end * rate)].copy()
|
||
assert len(clip) > rate * .2, line["id"]
|
||
fade = min(round(.005 * rate), len(clip) // 4)
|
||
for i in range(fade):
|
||
clip[i] *= i / fade
|
||
clip[-i - 1] *= i / fade
|
||
peak = max(abs(clip))
|
||
if peak > .98:
|
||
clip *= .98 / peak
|
||
target = output_path / (line["id"] + ".wav")
|
||
sf.write(target, clip, rate, subtype="PCM_16")
|
||
report.append(dict(id=line["id"], text=line["text"], start=round(start, 3),
|
||
end=round(end, 3), duration=round(len(clip) / rate, 3),
|
||
reference_words=b - a, matched_words=b - a, file=target.name))
|
||
(output_path / "alignment.json").write_text(json.dumps(report, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
|
||
print(f"Exported {len(report)} clips; every script word aligned")
|
||
|
||
if __name__ == "__main__":
|
||
main()
|