legco-speech / scripts /segment_audio.py
laubonghaudoi's picture
Super squash: clean up history after removing duplicate folders
4f81a3b
"""Segment full meeting audio into sentence-level clips and write Parquet shards.
For each meeting in metadata.csv, this script:
1. Parses the SRT file to get (start, end, text) segments.
2. Uses ffmpeg to extract each segment from the opus file (stream copy, no re-encoding).
3. Batches segments and writes Parquet shards with embedded audio bytes.
Usage:
python -m scripts.segment_audio [--workers N] [--shard-size N] [--out-dir DIR]
"""
import argparse
import collections
import csv
import itertools
import re
import subprocess
from concurrent.futures import FIRST_COMPLETED, ProcessPoolExecutor, wait
from dataclasses import dataclass
from pathlib import Path
from rich.console import Console
from rich.progress import (
BarColumn,
MofNCompleteColumn,
Progress,
SpinnerColumn,
TextColumn,
TimeElapsedColumn,
TimeRemainingColumn,
)
REPO_ROOT = Path(__file__).resolve().parent.parent
DEFAULT_SHARD_SIZE = 5000
DEFAULT_WORKERS = 4
console = Console()
@dataclass
class SrtSegment:
index: int
start_seconds: float
end_seconds: float
text: str
def _ts_to_seconds(ts: str) -> float:
"""Convert SRT timestamp (HH:MM:SS,mmm) to seconds."""
h, m, rest = ts.split(":")
s, ms = rest.split(",")
return int(h) * 3600 + int(m) * 60 + int(s) + int(ms) / 1000
def parse_srt(srt_path: Path) -> list[SrtSegment]:
"""Parse an SRT file into a list of timed segments."""
try:
content = srt_path.read_text(encoding="utf-8")
except FileNotFoundError:
return []
segments: list[SrtSegment] = []
blocks = re.split(r"\n\s*\n", content.strip())
for block in blocks:
lines = block.strip().split("\n")
if len(lines) < 3:
continue
try:
idx = int(lines[0].strip())
except ValueError:
continue
ts_match = re.match(
r"(\d{2}:\d{2}:\d{2},\d{3})\s*-->\s*(\d{2}:\d{2}:\d{2},\d{3})",
lines[1].strip(),
)
if not ts_match:
continue
start = _ts_to_seconds(ts_match.group(1))
end = _ts_to_seconds(ts_match.group(2))
text = " ".join(l.strip() for l in lines[2:] if l.strip())
if not text or end <= start:
continue
segments.append(SrtSegment(index=idx, start_seconds=start, end_seconds=end, text=text))
return segments
def extract_segment_audio(opus_path: Path, start: float, duration: float) -> bytes | None:
"""Extract a segment from an opus file using ffmpeg stream copy.
Returns the raw OGG/Opus bytes, or None on failure.
"""
cmd = [
"ffmpeg",
"-v", "error",
"-ss", f"{start:.3f}",
"-i", str(opus_path),
"-t", f"{duration:.3f}",
"-c", "copy",
"-f", "ogg",
"pipe:1",
]
try:
result = subprocess.run(cmd, capture_output=True, timeout=30)
if result.returncode != 0:
return None
if len(result.stdout) < 100:
return None
return result.stdout
except (subprocess.TimeoutExpired, OSError):
return None
def process_meeting(row: dict) -> list[dict]:
"""Process a single meeting: parse SRT, extract all audio segments."""
video_id = row["id"]
opus_path = REPO_ROOT / row["audio"]
srt_path = REPO_ROOT / row["subtitles"]
if not opus_path.exists():
return []
segments = parse_srt(srt_path)
if not segments:
return []
results = []
for seg in segments:
duration = seg.end_seconds - seg.start_seconds
audio_bytes = extract_segment_audio(opus_path, seg.start_seconds, duration)
if audio_bytes is None:
continue
results.append({
"video_id": video_id,
"segment_id": seg.index,
"audio": {"bytes": audio_bytes, "path": f"{video_id}_{seg.index:05d}.opus"},
"text": seg.text,
"start_time": round(seg.start_seconds, 3),
"end_time": round(seg.end_seconds, 3),
"duration": round(duration, 3),
})
return results
def write_shard(segments: list[dict], shard_idx: int, out_dir: Path) -> Path:
"""Write a list of segment dicts as a Parquet shard with Audio feature."""
from datasets import Audio, Dataset, Features, Value
features = Features({
"video_id": Value("string"),
"segment_id": Value("int32"),
"audio": Audio(),
"text": Value("string"),
"start_time": Value("float64"),
"end_time": Value("float64"),
"duration": Value("float64"),
})
ds = Dataset.from_dict(
{k: [s[k] for s in segments] for k in segments[0]},
features=features,
)
path = out_dir / f"train-{shard_idx:05d}.parquet"
ds.to_parquet(path)
return path
def _flush_buffer(buffer: collections.deque, shard_size: int, shard_idx: int,
out_dir: Path, *, force: bool = False) -> int:
"""Write complete shards from buffer. Returns updated shard_idx."""
while len(buffer) >= shard_size or (force and buffer):
n = min(shard_size, len(buffer))
batch = [buffer.popleft() for _ in range(n)]
shard_path = write_shard(batch, shard_idx, out_dir)
console.print(
f" Wrote shard {shard_idx} ({n} segments) -> {shard_path.name}"
)
del batch
shard_idx += 1
return shard_idx
def main() -> None:
parser = argparse.ArgumentParser(description="Segment audio and build Parquet shards")
parser.add_argument("--workers", type=int, default=DEFAULT_WORKERS,
help="Number of parallel workers (default: %(default)s)")
parser.add_argument("--shard-size", type=int, default=DEFAULT_SHARD_SIZE,
help="Segments per Parquet shard (default: %(default)s)")
parser.add_argument("--out-dir", type=Path, default=REPO_ROOT / "segmented",
help="Output directory for Parquet shards")
args = parser.parse_args()
args.out_dir.mkdir(parents=True, exist_ok=True)
src = REPO_ROOT / "metadata.csv"
with open(src, encoding="utf-8", newline="") as f:
reader = csv.DictReader(f)
rows = list(reader)
console.print(f"Processing {len(rows)} meetings with {args.workers} workers")
console.print(f"Shard size: {args.shard_size} segments")
console.print(f"Output: {args.out_dir}")
buffer: collections.deque[dict] = collections.deque()
shard_idx = 0
total_segments = 0
errors = 0
meetings_done = 0
max_in_flight = args.workers * 2
progress = Progress(
SpinnerColumn(),
TextColumn("[progress.description]{task.description}"),
BarColumn(),
MofNCompleteColumn(),
TimeElapsedColumn(),
TimeRemainingColumn(),
console=console,
)
with progress:
task = progress.add_task("Meetings processed", total=len(rows))
rows_iter = iter(rows)
with ProcessPoolExecutor(max_workers=args.workers) as pool:
active: dict = {}
# Seed the pool with an initial batch of work
for row in itertools.islice(rows_iter, max_in_flight):
f = pool.submit(process_meeting, row)
active[f] = row["id"]
while active:
done, _ = wait(active, return_when=FIRST_COMPLETED)
for future in done:
video_id = active.pop(future)
try:
segments = future.result()
buffer.extend(segments)
total_segments += len(segments)
del segments
except Exception as e:
errors += 1
console.print(f"[red]Error processing {video_id}: {e}[/red]")
# Release the Future's internal result reference
future._result = None
meetings_done += 1
progress.advance(task)
# Submit next meeting to keep the pool fed
row = next(rows_iter, None)
if row is not None:
f = pool.submit(process_meeting, row)
active[f] = row["id"]
# Flush complete shards after processing each batch of done futures
shard_idx = _flush_buffer(
buffer, args.shard_size, shard_idx, args.out_dir
)
# Flush remaining segments
shard_idx = _flush_buffer(
buffer, args.shard_size, shard_idx, args.out_dir, force=True
)
console.print(f"\n[bold green]Done![/bold green]")
console.print(f" Total segments: {total_segments}")
console.print(f" Total shards: {shard_idx}")
console.print(f" Errors: {errors}")
if __name__ == "__main__":
main()