Dual-channel audio isolation mechanics

Accurate speech transcription breaks down when host speech and participant speech collide in a single mono channel. Oyma maintains strict dual-channel audio separation throughout its capture and inference pipeline.

The dual-channel signal flow

When an audio session begins, Oyma instantiates two independent CoreAudio input taps:

1. Tap A (Microphone Input): Reads raw PCM samples from your configured hardware microphone (Internal mic, USB microphone, or AirPods).
2. Tap B (ScreenCaptureKit Loopback): Reads digital audio output generated by remote participants inside Zoom, Meet, Teams, or Slack.

┌──────────────────┐               ┌──────────────────┐
│   Local Mic      │               │  Call Loopback   │
│  (48kHz, Mono)   │               │  (48kHz, Mono)   │
└────────┬─────────┘               └────────┬─────────┘
         │                                  │
         ▼                                  ▼
┌─────────────────────────────────────────────────────┐
│             Interleaved Stereo PCM Buffer           │
│         [Left: Channel 0]    [Right: Channel 1]     │
└──────────────────────────┬──────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────┐
│             Voice Activity Detector (VAD)           │
│           Independent Turn-Taking Detection         │
└─────────────────────────────────────────────────────┘

Benefits of independent channel processing

• Zero Acoustic Bleed: Remote participants can speak over the host without voice waveform mixing.
• Independent VAD (Voice Activity Detection): Background coughing or keyboard typing on your mic does not interrupt participant transcription.
• Accurate Speaker Attribution: Channel index deterministically tags host vs remote speakers with 100% certainty.

Buffer format specifications

• Sample Rate: 16,000 Hz (native input rate required by Whisper). CoreAudio handles hardware sample rate conversion from 44.1 kHz or 48 kHz.
• Bit Depth: 32-bit floating point internally, converted to 16-bit PCM for Whisper inference.
• Channels: 2 discrete logical channels.

Related guides

• Pipeline architecture in System audio pipeline and ScreenCaptureKit.
• Hardware configuration in Headset selection and echo cancellation.
• Speaker assignment in Speaker attribution and turn-taking notes.