🎙️ Algorithmic Voice Detection & Adaptive Recorder (No AI / Pure DSP)

A high-precision, low-latency Voice Activity Detection (VAD) and automated audio recording engine built entirely on Digital Signal Processing (DSP) principles without external machine learning, cloud dependencies, or neural network overhead.

---

🔬 Mathematical Theory & Signal Processing Pipeline

[Raw Audio Stream] (16kHz, 16-bit PCM)
        │
        ▼
[30ms Window Partitioning] (480 samples/frame)
        │
        ▼
[Fast Fourier Transform (FFT)] ──► Real-to-Complex Frequency Domain (np.fft.rfft)
        │
        ▼
[Bandpass Masking] ──────────────► Zero-out frequencies < 300Hz and > 3400Hz (Vocal Tract)
        │
        ▼
[Inverse FFT (np.fft.irfft)] ────► Cleaned time-domain acoustic waveform
        │
        ▼
[RMS Energy Calculation] ────────► E_rms = sqrt( (1/N) * sum( x[i]^2 ) )
        │
        ▼
[Adaptive Silence Decision Engine] ─► Stops recording after trailing silence threshold

Why Algorithmic DSP over Deep Learning?

• Zero Latency: Operates in real-time per 30ms window with sub-millisecond execution overhead.
• Minimal Footprint: No PyTorch, TensorFlow, or ONNX runtimes. Consumes less than 25MB RAM.
• Bandpass Isolation: Suppresses low-frequency mechanical rumble (< 300Hz) and high-frequency hiss (> 3400Hz), isolating the fundamental formants of human speech.
• Dynamic Ambient Calibration: Automatically measures background acoustic levels on startup to adapt the speech threshold.

---

🚀 Quick Start

1. Installation

git clone https://github.com/Gaoc3/VoiceDetection.git
cd VoiceDetection
pip install -r requirements.txt

2. Usage

Command Line Interface (CLI):

python voice_detector.py --output my_recording.wav --silence 2.0

As a Python Module:

from voice_detector import AcousticVoiceDetector

detector = AcousticVoiceDetector(
    sample_rate=16000,
    silence_sec=2.3
)

# Automatically calibrates noise and records until silence
detector.record_until_silence("meeting_note.wav")

---

⚙️ Configurable Parameters

Parameter	Type	Default	Description
sample_rate	int	16000	Audio sampling rate in Hertz.
frame_ms	int	30	Analysis window chunk size in milliseconds.
low_cutoff_hz	float	300.0	Lower bandpass frequency bound (human voice band).
high_cutoff_hz	float	3400.0	Upper bandpass frequency bound.
silence_sec	float	2.3	Trailing silence duration before stopping recording.
rms_threshold	float	500.0	Base speech sensitivity threshold (dynamically calibrated).

---

👨‍💻 Author

• Hussain Ibrahim Ahmed (Gaoc3)
  Systems & Backend Software Engineer | Distributed Infrastructure
  LinkedIn • GitHub