Oyma v0.3.0: CoreML Whisper Turbo & Multilingual Models

Release Highlights: Apple Silicon Hardware Acceleration

Running large automatic speech recognition models locally requires substantial compute power. Version 0.3.0 introduces CoreML acceleration for Whisper models on Apple Silicon.

Performance Benchmarks

Neural Engine Offloading

By compiling Whisper model graphs to Apple CoreML format, inference runs concurrently on the GPU and Apple Neural Engine. This results in minimal battery consumption on MacBook Pro and MacBook Air hardware, allowing users to transcribe multi-hour meetings while unplugged.

File Transcription Mode

In addition to live meeting capture, users can now drag and drop pre-recorded audio files (MP3, WAV, M4A, MP4) directly into the app for high-speed offline transcription.

Quantization Benchmarks & Memory Architecture

We evaluated the quantized CoreML execution graph across multiple Apple Silicon architectures:

• Whisper Large v3 Turbo FP16 requires only 1,120MB of unified memory, allowing seamless execution alongside memory-intensive IDEs.
• Audio preprocessing utilizes vDSP hardware acceleration for ultra-fast Mel-spectrogram calculation.
• Verified support for 13 global languages including English, Spanish, German, French, Japanese, and Mandarin.

Check our audio and video transcription guide, select the right model with our Whisper model picker, inspect live transcription features, and read the transcription models documentation.