An hour of audio transcribed in about 13s on MacBook Air

On-device audio transcription performance on consumer laptops improves significantly.

Google DeepMind has developed a speech-to-text model called Embedding Gemma Transcribe, derived from Google's EmbeddingGemma 2 multimodal model. This new model can transcribe an hour of audio in approximately 13 seconds when running on a base MacBook Air, processing entirely on the device. The system utilizes EmbeddingGemma 2's audio encoder and a small, newly trained "reader" model that maps audio slices to the closest English words. An additional tiny model handles punctuation and capitalization. The reader model was trained in under two hours on eight H100 GPUs using openly licensed English speech data. The base EmbeddingGemma 2 model itself was not retrained for this task. This independent project, released under Apache 2.0, runs on the Mac's Neural Engine, ensuring audio privacy as it never leaves the machine. Its current word error rate is around 5%, with expectations of improvement.

The innovation lies in repurposing a multimodal embedding model for transcription, demonstrating a novel approach to ASR that leverages pre-existing model capabilities rather than traditional end-to-end training. This method bypasses extensive retraining of the core model, instead focusing on a small, efficiently trained "reader" component. The speed achieved on consumer hardware suggests a potential shift towards more capable, privacy-preserving AI applications running locally, reducing reliance on cloud-based services for such tasks.

This development could impact the competitive landscape for speech-to-text services, particularly for applications prioritizing on-device processing and user privacy. While the current word error rate is slightly higher than leading cloud models, its local processing advantage and speed on accessible hardware make it a compelling option for certain use cases. Future improvements to the reader model will determine its broader market viability against established players.