✓Direct Answer

Oyma runs an optimized CoreML implementation of OpenAI's Whisper model directly on your Mac's Apple Silicon Neural Engine. Speech is transcribed locally in real time without sending audio data across the internet.

Machine Learning Speech Recognition at the Edge

OpenAI’s Whisper revolutionized automatic speech recognition (ASR) by delivering human-level accuracy across diverse accents, background noise environments, and technical terminology. However, running Whisper through cloud APIs costs money per minute and transmits confidential voice data to remote servers.

Oyma brings Whisper directly onto your Mac hardware, utilizing Apple’s specialized Apple Neural Engine (ANE).

The CoreML Neural Engine Pipeline

When speech audio enters Oyma‘s transcription pipeline:

  1. Audio Chunking: AVFoundation processes linear PCM audio into 30-second log-mel spectrogram frames.
  2. CoreML Hardware Dispatch: The spectrogram tensors are fed into an optimized CoreML graph compiled specifically for Apple Silicon hardware.
  3. Neural Engine Execution: The 16-core Apple Neural Engine executes matrix multiplications with minimal CPU and GPU power draw.
  4. Token Decoding: The beam-search decoder outputs punctuated, time-aligned text segments directly into your active Markdown note.
[Raw Audio Stream]
        |
        v
[30s Mel Spectrogram]
        |
        v
[CoreML Graph] ---> [Apple Silicon Neural Engine (ANE)]
        |
        v
[Punctuated CommonMark Transcript with Timestamps]

Real-Time vs. Post-Call Transcription

  • Real-Time Streaming Mode: Generates live subtitles as participants speak, displaying a running transcript alongside your note.
  • Batch Processing Mode: Transcribes audio at high speed after the meeting concludes. On an M3 Pro Mac, a 60-minute meeting is transcribed in approximately 2.5 minutes using Whisper Base.

Because transcription occurs entirely on-device, processing latency is virtually zero and your conversations remain strictly confidential.

Technical Architecture of CoreML Whisper

Apple Silicon’s unified architecture enables remarkable automatic speech recognition performance:

  • 30-Second Log-Mel Spectrograms: AVFoundation audio buffers are preprocessed into 80-channel log-mel spectrogram tensors.
  • Transformer Encoder Execution: The spectrogram is processed by Whisper’s encoder stack running on the Apple Neural Engine, producing audio feature representations.
  • Autoregressive Text Decoder: The decoder generates token probabilities with beam search decoding, adding accurate punctuation, capitalization, and timestamps.
  • Zero Thermal Penalty: Because the Apple Neural Engine handles the matrix multiplications, the main CPU and GPU cores remain cool, preserving battery life and system responsiveness.

Uninterrupted Neural Engine Throughput on Apple Silicon

By running Whisper on the dedicated Apple Neural Engine, Oyma leaves your Mac’s CPU and GPU cores completely free for video conferencing and text editing. You experience zero UI stutter or dropped video frames during active transcription.

The CoreML graph is compiled to utilize 16-bit floating point precision (FP16), cutting memory bandwidth requirements in half while preserving word accuracy identical to full 32-bit floating point models.

Related Questions & Architecture Guides