Research areas

Overview

Research areas

Speech and voice

In a voice conversation, a second of silence feels broken. We build voice systems where recognition, reasoning and speech synthesis work as one pipeline, tuned for low latency, natural turn-taking and the accents and noise of real calls.

  1. audio in

    phone · browser

  2. recognise

    streaming ASR

  3. turn-taking

    pauses · interruptions

  4. respond

    LLM · tools

  5. speak

    streaming TTS

Fig. 1A streaming voice pipeline where every stage adds to the time before a reply

Typical problems

  • Replies arrive too slowly for the conversation to feel natural
  • The assistant talks over users, or waits too long after they finish
  • Recognition breaks on accents, background noise or domain terms
  • Calls need to trigger real actions, like bookings or account changes

What you get

  • A streaming voice pipeline with a measured latency budget per stage
  • Turn-taking and interruption handling tested on real recordings
  • Vocabulary and prompt tuning for your domain and languages
  • Call-level evaluation and monitoring of task success

How we approach it

We treat latency as a budget and measure where every millisecond goes, from audio capture to the first syllable of the reply. Each stage streams into the next, so the system starts speaking before it has finished thinking.

Turn-taking gets as much attention as the model itself. We test how the system handles pauses, interruptions and crosstalk on recordings that sound like your real calls, not just clean studio audio.