Voice-Activated Auto-Record (VAD)

1. Introduction

Voice-Activated Auto-Record (VAD), referred to in the codebase as “pre-recording audio activity detection,” is a mechanism designed to gate the recording process based on the presence of audio signals. Its primary role is to prevent the system from capturing and processing silent or empty audio buffers, which would otherwise waste computational resources during the transcription phase. The system operates by monitoring input levels and blocking capture triggers if the signal does not meet specific activity thresholds.

Sources: CHANGELOG.md, README.md

2. Audio Monitoring and Signal Detection

The VAD mechanism is integrated into the AudioEngine and the UI’s VUMeter components. The system calculates decibel (dB) levels and peak amplitudes from the JACK input ports to determine if a “thought” is actually being spoken.

2.1 Metering and Thresholds

The AudioEngine maintains real-time peak and dB calculations for each input channel. These values are used by the UI to provide visual feedback and by the recording logic to validate if a capture should proceed.

Sources: src/omega13/audio.py:#L46-L47, src/omega13/ui.py:#L14-L32

2.2 Structural Flow of Signal Validation

The system utilizes a rolling NumPy ring buffer to store the last 13 seconds of audio. When a capture is triggered (via global hotkey or CLI), the system checks the buffer state. If no signal is detected—a state described in the documentation as “Capture Blocked - No Input Signal”—the recording process is aborted.

Note: The “Capture Blocked” state is an explicit safety check to ensure the local AI model (Whisper) is not invoked on silence.

Sources: src/omega13/audio.py:#L66-L95, README.md

3. Implementation Details

3.1 Buffer Management

The AudioEngine handles the 13-second “retroactive” window. The VAD logic implicitly relies on the buffer_filled flag and the write_ptr to determine if there is sufficient data to even evaluate for activity.

# src/omega13/audio.py:#L79-L85

if self.buffer_filled:
    part_old = self.ring_buffer[self.write_ptr:]
    part_new = self.ring_buffer[:self.write_ptr]
    past_data = np.concatenate((part_old, part_new))
else:
    past_data = self.ring_buffer[:self.write_ptr].copy()

Sources: src/omega13/audio.py:#L79-L85

3.2 Integration with UI

The Omega13App coordinates the VAD feedback loop. If the AudioEngine reports peaks below the threshold, the status-bar in the TUI remains IDLE or displays a warning, preventing the transition to the RECORDING state (marked by a red UI change).

Sources: src/omega13/app.py:#L45-L55, src/omega13/app.py:#L174-L185

4. Observed Structural Inconsistencies

The system presents a curious contradiction in its “always listening” philosophy. While it maintains a constant 13-second buffer, the actual VAD logic appears to be a “gatekeeper” at the moment of trigger rather than a continuous start/stop mechanism. This means the system is fucking listening all the time, but it only decides if that listening was “worth it” once you’ve already hit the button.

Furthermore, the TranscriptionService is strictly dependent on the output of this gated recording. If the VAD logic fails to block a silent recording, the system will proceed to send an empty .wav file to the whisper-server, leading to unnecessary HTTP overhead and potential “processing” of background noise.

Sources: src/omega13/transcription.py:#L57-L70, README.md

5. Conclusion

The Voice-Activated Auto-Record (VAD) mechanism in Omega-13 serves as a critical filter between the continuous audio capture of the JACK client and the resource-heavy AI transcription backend. By leveraging real-time peak detection and decibel monitoring, it ensures that only meaningful audio segments are promoted from the temporary ring buffer to permanent session storage. Its structural significance lies in its role as a resource-saver for local GPU/CPU inference.