CLIEncoders · AI & Agents
Voice AI and speech system development
Voice is the hardest interface to get right, because the tolerance for delay is about a third of a second and there is no way to skim what was said.
Latency is the whole problem
A text assistant can take three seconds to answer and feel fine. A voice assistant taking three seconds feels broken — callers start talking over it, and the conversation degrades from there.
So the engineering is mostly about the pipeline: streaming transcription rather than waiting for the caller to finish, beginning to generate before the full transcript arrives, and streaming synthesis so the first syllable leaves early. Every stage is chosen for latency, which sometimes means a smaller, faster model beats a more capable slow one.
Interruption and turn-taking
People interrupt. They pause mid-sentence, they say 'um', they answer a question you have not finished asking. Handling this — barge-in detection, knowing when a pause is a finished turn rather than thinking time — is what separates a system people tolerate from one that infuriates them.
We also plan for the recogniser being wrong. Names, addresses and reference numbers are where transcription fails most, so those get confirmation rather than being acted on directly.
What we build
Inbound and outbound voice agents integrated with telephony, real-time and batch transcription with speaker separation, meeting and call summarisation, IVR replacement that lets callers say what they want instead of navigating a menu tree, and voice interfaces embedded in devices and apps.
Questions
What people ask before starting
How accurate is transcription on accented speech?
Good and improving, but not uniform — accuracy varies by accent, background noise and audio quality, and phone audio is narrowband and unkind. For your specific callers we test on real recordings before promising anything, because a benchmark figure from clean studio audio tells you very little about a call centre.
Will callers know they are talking to an AI?
They should, and we build it that way. Disclosure at the start of the call is a legal requirement in a growing number of jurisdictions and the right default regardless. Voices are also plainly synthetic rather than impersonating a specific real person — cloning someone's voice without their explicit consent is not something we will build.
Can it transfer to a human?
Yes, and it needs to. Transfer triggers on explicit request, on detected frustration, and on any topic you have marked as human-only. The human receives the transcript and context rather than starting cold, which is the difference between a transfer and an abandonment.
What does voice cost per minute?
There are three meters running — transcription, the language model, and synthesis — plus telephony. It is materially more expensive per interaction than text, and worth modelling against your call volume before committing. We will do that up front, and if the numbers do not work we will say so.
Related
Where this usually connects
Tell us what you are building
One technical call is usually enough to tell you whether this is straightforward, genuinely hard, or the wrong approach entirely. We would rather say so early than quote for the wrong thing.
Start the conversation