What's this about?
Speko, a Y Combinator S26 company, launched describing itself as "OpenRouter for Voice AI" — a single routing layer across many voice models. The timing is conspicuous: Stripe agreed to buy the original OpenRouter for over $7 billion the day before.
What exactly is voice AI routing, and why is it its own problem?
Voice systems chain several models: speech recognition turning audio into text, a language model deciding what to say, and speech synthesis producing the reply. Each stage has many providers with different accuracy, latency, language coverage and price. Routing text models is mostly about quality against cost; routing voice adds a hard real-time constraint, because a delay that's imperceptible in a chat window is a broken conversation in a phone call. That latency budget is why voice routing is a genuinely different engineering problem rather than the same product with audio attached.
Why is this worth noticing?
- "X for Y" positioning usually invites skepticism, but a $7B validation of the category the day before makes it an unusually well-timed pitch.
- It marks routing maturing from a convenience into an infrastructure layer that gets rebuilt per modality — text yesterday, voice today, presumably video next.
- The category's strategic question applies here too: a router sees which model is chosen for which job, which is valuable data, and which is precisely why the last one sold for billions.