segment_vad.py

#!/usr/bin/env python3
import argparse
import math
from pathlib import Path
from typing import List, Tuple

import numpy as np
import soundfile as sf
import webrtcvad
from tqdm import tqdm


def read_audio(path: Path) -> Tuple[np.ndarray, int]:
	data, sr = sf.read(str(path))
	if data.ndim == 2:
		data = data.mean(axis=1)
	return data.astype(np.float32), sr


def write_audio(path: Path, data: np.ndarray, sr: int) -> None:
	path.parent.mkdir(parents=True, exist_ok=True)
	sf.write(str(path), data, sr)


def frame_generator(audio: np.ndarray, sample_rate: int, frame_ms: int) -> List[np.ndarray]:
	samples_per_frame = int(sample_rate * frame_ms / 1000)
	num_frames = int(math.ceil(len(audio) / samples_per_frame))
	frames = []
	for i in range(num_frames):
		start = i * samples_per_frame
		end = min((i + 1) * samples_per_frame, len(audio))
		frame = audio[start:end]
		if len(frame) < samples_per_frame:
			pad = np.zeros(samples_per_frame - len(frame), dtype=audio.dtype)
			frame = np.concatenate([frame, pad], axis=0)
		frames.append(frame)
	return frames


def vad_segments(audio: np.ndarray, sample_rate: int, aggressiveness: int = 2, frame_ms: int = 30,
                 min_speech_ms: int = 300, max_silence_ms: int = 300) -> List[Tuple[int, int]]:
	assert frame_ms in (10, 20, 30), "WebRTC VAD supports 10/20/30 ms"
	vad = webrtcvad.Vad(aggressiveness)
	frames = frame_generator(audio, sample_rate, frame_ms)
	is_speech = []
	for frame in frames:
		int16 = np.clip(frame * 32768.0, -32768, 32767).astype(np.int16).tobytes()
		is_speech.append(vad.is_speech(int16, sample_rate))
	ms_per_frame = frame_ms
	segments = []
	start = None
	run_silence = 0
	for i, speech in enumerate(is_speech):
		if speech:
			if start is None:
				start = i
			run_silence = 0
		else:
			if start is not None:
				run_silence += ms_per_frame
				if run_silence >= max_silence_ms:
					end = i
					duration_ms = (end - start) * ms_per_frame
					if duration_ms >= min_speech_ms:
						segments.append((start, end))
					start = None
					run_silence = 0
	# tail
	if start is not None:
		segments.append((start, len(is_speech)))
	# convert to sample indices
	spf = int(sample_rate * frame_ms / 1000)
	sample_segments = [(s * spf, e * spf) for s, e in segments]
	return sample_segments


def hard_cap_segments(segments: List[Tuple[int, int]], max_seconds: float, sample_rate: int) -> List[Tuple[int, int]]:
	max_len = int(max_seconds * sample_rate)
	capped = []
	for s, e in segments:
		length = e - s
		if length <= max_len:
			capped.append((s, e))
		else:
			num = math.ceil(length / max_len)
			for i in range(num):
				cs = s + i * max_len
				ce = min(s + (i + 1) * max_len, e)
				if ce - cs > int(0.2 * sample_rate):  # drop ultra-short tails
					capped.append((cs, ce))
	return capped


def process_file(in_path: Path, out_dir: Path, max_seconds: float) -> int:
	audio, sr = read_audio(in_path)
	if sr != 16000:
		# Expect normalized inputs; skip others
		return 0
	segments = vad_segments(audio, sr, aggressiveness=2, frame_ms=30, min_speech_ms=250, max_silence_ms=300)
	segments = hard_cap_segments(segments, max_seconds, sr)
	stem = in_path.stem
	for idx, (s, e) in enumerate(segments):
		seg = audio[s:e]
		out_path = out_dir / f"{stem}_seg{idx:04d}.wav"
		write_audio(out_path, seg, sr)
	return len(segments)


def main() -> None:
	parser = argparse.ArgumentParser(description="VAD-based segmentation to <= max seconds, expects 16kHz mono WAV inputs.")
	parser.add_argument("--indir", required=True, help="Input audio root")
	parser.add_argument("--outdir", required=True, help="Output audio root (can be same as input subfolders)")
	parser.add_argument("--max-seconds", type=float, default=30.0, help="Maximum segment length in seconds")
	args = parser.parse_args()
	inroot = Path(args.indir)
	outroot = Path(args.outdir)
	wavs = list(inroot.rglob("*.16k.wav"))
	if not wavs:
		# fallback: all wavs
		wavs = list(inroot.rglob("*.wav"))
	total_segments = 0
	for f in tqdm(wavs, desc="Segmenting"):
		out_dir = f.parent
		total_segments += process_file(f, out_dir, args.max_seconds)
	print(f"Total segments written: {total_segments}")


if __name__ == "__main__":
	main()