From the team
Notes on voice AI, language technology, and building for the next billion users.
You can call Aika on WhatsApp now
Aika now answers WhatsApp calls and calls people back on WhatsApp. Why we built it, what changes in the audio, and the two bugs that only a real call could show us.
Read the postWho is on the line? Age checks on a shared family phone
In India one phone often serves a whole family, so the number is not the person. How we check for a child on every call, and what a voice can and can't tell.
The rig is our eval
Unit tests can't tell you whether a voice assistant handled a call well. We replay whole conversations, and every call that goes wrong becomes a new scenario.
Cold and warm: prefix caching on a self-hosted voice model
Our self-hosted model took nearly two seconds to start answering, and we were sure caching was broken. It wasn't. What we learned about the first call on a warm GPU.
Let the voice model be a voice model
We used to give our conversation model a menu of tools and let it decide, extract and save. Now code decides and acts, and the model just speaks. Why we moved.
Why wideband phone calls are out of reach in India, for now
Wideband audio should help speech recognition on phone calls, so we tried to get it for our callers. Every hop of an Indian call has to agree, and most don't.
How a phone call reaches a voice agent
Before a speech model hears a caller, their voice crosses a handset, two operators, a carrier trunk and a jitter buffer. What each hop does to the audio.
We replaced our decision stack with Jev. It worked.
How Aika replaced embedding and LLM decisions with Jev in live voice calls, and what batching, prefetching and checkout taught us.
The wrong kind of conversation: why helpful assistants feel cold
An assistant that answers a complaint with correct advice is in the wrong conversation. A 1981 paper on advice-rejection explains why that feels cold.
Performance is accessibility
Most of the world is still on cheap phones and slow networks. Why page weight is an accessibility problem, and what it means to design for the 2G web.
On Arm CPUs, 8-bit can beat 4-bit
A counterintuitive result for running LLMs on Arm servers: 8-bit often does prefill faster than 4-bit, and most builds leave a free int8 speedup switched off.
Knowing when to talk: turn-taking and the Indian-language data gap
Voice agents interrupt people or leave them hanging. For Indian languages, the fix isn't a bigger model. It's the conversational data nobody collected.