SEAD - Sound Event Activity Detection

Sound Event Activity Detection (SEAD). Detects speech, music, and other sounds in audio with offline and streaming support.

Installation

pip install sead

For microphone streaming, also install the optional sounddevice dependency:

pip install sead[dev]

For GPU inference (optional):

pip install sead[onnx-gpu]

Usage

CLI

# Offline: process a WAV file
sead --audio audio_16khz.wav

# Or via Python module
python -m sead.cli --audio audio_16khz.wav

# Synthetic streaming: feed file as streaming input (emits start/end events)
sead --stream-file audio_16khz.wav

# Compare streaming vs offline
sead --stream-file audio_16khz.wav --compare

# Live microphone streaming (requires sounddevice)
sead --stream

# Save debug chunks for each segment
sead --audio audio_16khz.wav --debug-dir debug_segments

Python API

from sead import DEFAULT_MODEL_PATH, SEADDetector, SEADIterator, Segment
from pathlib import Path

# Offline
detector = SEADDetector(DEFAULT_MODEL_PATH)
segments = detector.process_file(Path("audio.wav"))
for s in segments:
    print(s)  # [start_time, end_time, label, confidence]

# Streaming (incremental events)
iterator = SEADIterator(detector)
for chunk in audio_chunks:
    for e in iterator(chunk):
        print(e)  # {'start': t, 'label': str} or {'end': t, 'label': str, 'confidence': float}
for e in iterator.flush():
    print(e)

Output

• Segments: [start_time, end_time, label, confidence] with labels speech, music, others
• Events (streaming): {'start': t, 'label': str} on onset, {'end': t, 'label': str, 'confidence': float} on offset

Acknowledgement

SEAD uses YamNet by Qualcomm, an audio event classifier trained on the AudioSet dataset.