Oyma v0.10.0: Streaming Whisper & Dual-Channel VAD

Release Highlights: Real-Time Acoustic Processing

Version 0.10.0 brings significant enhancements to our audio ingestion and machine learning pipeline, introducing low-latency streaming transcription and intelligent voice activity filtering on Apple Silicon.

Key Technical Advancements

Streaming Voice Activity Detection (VAD)

We incorporated an on-device Silero VAD model operating on 30ms audio windows. The detector accurately filters background keyboard typing, fan noise, and ambient cafeteria sounds before feeding audio frames into Whisper, reducing inference power consumption by 34% during meeting pauses.

Lock-Free Audio Memory Pool

To ensure that CoreAudio callbacks are never blocked by garbage collection or filesystem I/O, the audio capture subsystem now uses lock-free circular ring buffers allocated directly in pinned memory. This guarantees zero frame drops during heavy CPU load.

Interactive Dialogue Timestamps

Transcripts generated during live meetings now feature interactive timestamp badges ([00:14:22]). Clicking any timestamp seeks the embedded audio player to that exact moment in the recording for immediate verification. Learn more about our live transcription capabilities.

Technical Verification & Benchmarks

During sustained 90-minute benchmark sessions across Apple Silicon hardware (M1 Max, M2 Pro, and M3 Max), the streaming pipeline demonstrated:

• End-to-end transcription latency averaging 1,180ms from spoken word to screen render.
• Peak RAM consumption under 1.25GB during active transcription.
• 0% dropped audio packets across all tested sample rates (16kHz to 48kHz).

Review our speaker talk time analyzer tool, inspect the transcription models guide, or discover local AI with Ollama.