MacBook Air에서 1시간 분량 오디오를 약 13초 만에 전사

Google's EmbeddingGemma 2, originally a multimodal embedding model, has been adapted into a speech-to-text system that runs entirely on-device. By leveraging the model's shared space for text and audio, the new tool reads words directly from the embedding space, achieving significant speed on Apple Silicon hardware.

How It Works

The adaptation preserves EmbeddingGemma 2's original audio encoder and adds a small, trained reader component:

• The reader maps short audio slices into the model's text space.
• The closest match from approximately 229,000 English words is selected to form the transcript.
• A secondary tiny model adds punctuation and capitalization.
• Only the reader was trained, requiring under two hours on 8×H100 GPUs using openly licensed English speech data.

Performance Benchmarks

The tool achieves 288× real-time speed on a base M3 MacBook Air (8 GB RAM), transcribing one hour of audio in 12.6 seconds. This performance utilizes the Mac's Neural Engine.

Comparative benchmarks provided in the source:

• Embedding Gemma Transcribe (M3 MacBook Air, Neural Engine): 288× (1 hour in 12.6 s)
• IBM Granite Speech 5.0 TurboCTC (M3 MacBook Air, GPU/MLX): ~150× (1 hour in 25.7 s)
• NVIDIA Parakeet TDT 0.6B v3 (M4 Pro, Neural Engine): 156×
• NVIDIA Parakeet TDT 0.6B v3 (M5 Pro, Neural Engine): 335×
• Phonon-2 (M5 Pro, Neural Engine): 480×

Accuracy and Availability

The current word error rate is approximately 5%, compared to 2–3% for leading dedicated speech models. The developers aim to close this gap in future iterations by improving the reader component.

The tool is available for free on macOS 14+ with Apple Silicon. It requires a 715 MB model download on first launch and processes audio locally, ensuring data never leaves the machine.