Whistle Packs Multilingual Speech-to-Text into 16.9 MB.

Whistle makes local voice interfaces substantially more practical for constrained devices, though its real-world accuracy still needs independent testing.

• –Runs offline across phones, wearables, robots, vehicles, browsers, and microcontrollers
• –Combines transcription with keyword biasing, silence handling, timestamps, and embeddings
• –Cactus reports faster inference and lower size than Whisper base on several benchmarks
• –Published results are company-reported, with Whisper still ahead on some datasets
• –Shared runtime enables a direct path from speech input to local structured tool calls