Small models, careful engineering.
Most phones and laptops have no powerful GPU. We build for the processor they do have, so more people can use AI without a fast connection or a new device.
The voice pipeline
- 01 Listen Voice activity detection finds where your sentence starts and ends, and ignores the app’s own voice coming back through the speaker.
- 02 Understand A compact speech-to-text model turns what you said into text without sending audio away.
- 03 Think A small language model, constrained to a fixed output format, writes the reply inside the phone’s memory limits.
- 04 Speak Text-to-speech starts with the first short phrase and keeps generating while it talks, so the reply begins sooner.
Quantised models
Smaller number formats, such as 8-bit weights, cut memory use so models fit on ordinary devices.
Streaming speech
Replies are split into natural phrases and spoken while the rest is still being generated.
Device profiles
Quick, Balanced and Expressive presets match voice quality to what each phone can handle.
Memory budgets
Short context windows and on-device storage keep the app inside mobile memory limits.
What runs where · KoalaSpeak
Conversation: listening, understanding, replying, speaking
ON DEVICE
Runs on the phone after the voice and language data has been downloaded.
Memory between conversations
ON DEVICE
Off by default. When switched on, it is stored on the device.
Downloading voice and language data
NEEDS INTERNET
Needed once per voice. Wi‑Fi recommended.
News radio
NEEDS INTERNET
Headlines are fetched from news sources.
Weather
NEEDS INTERNET
Forecasts are fetched for the city you choose.