v0.1.3 - ultra-lightweight package with pure-numpy frontend

pulsevad v0.1.3: ultra-lightweight release

this release slashes the package footprint from 69 packages down to only ~8 packages total, making pulsevad even lighter to install than Silero while maintaining bit-level numerical parity.

what changed:

• ultra-lightweight core dependencies:
  cut default runtime dependencies from 13 heavy libraries down to 4:

  dependencies = [
      "numpy>=1.24.0",
      "onnxruntime>=1.16.0",
      "scipy>=1.11.0",
      "soundfile>=0.12.1",
  ]

v0.1.2 - Updated metrics on various Indian languages

Full Changelog: v0.1.1...v0.1.2

v0.1.1 - Complete documentation & model usage guide

PulseVAD v0.1.1

This release bundles full PyPI project documentation, updated package metadata, and comprehensive usage guides for all pre-trained model flavors (ONNX, PyTorch TorchScript JIT, and embedded C header).

what's in this release

• Full pypi documentation: wired README.md into package metadata so the entire guide, benchmark tables, and code snippets render directly on PyPI.
• Bundled model matrix: wheels now explicitly bundle all 7 pre-trained model artifacts inside pulsevad/data/:
  • pulsevad_2.1k_int8.onnx (default 2.1 KB INT8 quantized model)
  • pulsevad_2.1k.onnx (FP32 ONNX)
  • pulsevad_2.1k.jit (TorchScript JIT for zero-ONNX PyTorch environments)
  • pulsevad_2.1k.pth (PyTorch state dict)
  • pulsevad_weights.h (standalone C header with INT8 arrays for microcontrollers)
  • pulsevad_teacher_81k.onnx & pulsevad_teacher_81k.pth (81k teacher baseline)
• Detailed usage guides: added copy-pasteable examples in README.md for:
  1. ONNX Runtime (INT8 vs FP32 vs Teacher)
  2. PyTorch TorchScript JIT
  3. Real-time streaming chunk-by-chunk processing (200 ms causal window)
  4. Audio file timestamp segmentation (get_speech_timestamps)
  5. Embedded C firmware integration on ARM Cortex-M, ESP32, and STM32
• Benchmarks:
  • Outperforms Silero v6 on windy audio (0.938 vs 0.934 AUC) and DNS synthetic noise (0.891 vs 0.873 AUC) at 145x smaller size.
  • Outperforms NVIDIA MarbleNet on clean speech (0.976 vs 0.970 AUC) and windy audio (0.938 vs 0.910 AUC) at 43x smaller size and 3.1x lower latency.
  • Validated across 10 global & Indian languages (FLEURS) with +0.029 higher macro average AUC over Silero in noise.

v0.1.0 - Initial release: 2.1k ultra-tiny VAD for edge deployment

PulseVAD v0.1.0

An ultra-compact, 2,118-parameter voice activity detector designed for microcontrollers and edge hardware (2.1 KB INT8). strictly causal with zero lookahead.

Highlights

• Ultra-tiny footprint: 2,118 parameters (2.1 KB as INT8, 8.5 KB as FP32).
• Drop-in c header: ships with pulsevad_weights.h containing self-contained INT8 weight arrays and scales for embedded C/C++ firmware.
• Silero-compatible python api:

  from pulsevad import load_pulsevad, read_audio, get_speech_timestamps
  model = load_pulsevad(onnx=True, quantized=True)
  wav = read_audio("speech.wav")
  timestamps = get_speech_timestamps(wav, model)