Speko Brings the OpenRouter Playbook to Voice AI
Launch HN today: Speko (YC S26), a router for voice AI stacks. It continuously benchmarks speech-to-text, LLM and text-to-speech models across languages and regions, then routes each request to the current best combination given your constraint, accuracy, latency, cost or balanced. The open-source gateway is free and self-hostable with your own keys; they charge for the hosted router and managed key consolidation. Usage has grown about 25 percent weekly since late June. speko.ai
The numbers that make the case: word error rates on medical terminology vary 8 to 19 percent across models, and their turn-taking detection claims 94 percent accuracy versus 46.9 for basic voice activity detection. Founder Bek spent four years building voice agents across Asia in ten-plus languages, and his diagnosis rings true, teams benchmark once at integration time and never again, because switching vendors means rework. Routing turns the switch into a config change. To their credit, the public benchmarks include cases where competitors beat Speko's own selections.
The timing is almost comically on-theme: the routing thesis arrives at the voice layer one day after Stripe reportedly agreed to buy OpenRouter for $7 billion plus. Every layer of the stack is growing a meter and a switchboard.
HN pushed back in two directions, and both are worth keeping. End-to-end speech models may kill the cascaded three-model architecture Speko optimizes (their counter: 95 percent of production systems are still cascaded), and local models may kill hosted routing entirely. Notice both objections are about whether the layer survives long-term, not whether it is valuable today. That was the OpenRouter objection too, and OpenRouter just sold for $7 billion.
← Back to all articles
The numbers that make the case: word error rates on medical terminology vary 8 to 19 percent across models, and their turn-taking detection claims 94 percent accuracy versus 46.9 for basic voice activity detection. Founder Bek spent four years building voice agents across Asia in ten-plus languages, and his diagnosis rings true, teams benchmark once at integration time and never again, because switching vendors means rework. Routing turns the switch into a config change. To their credit, the public benchmarks include cases where competitors beat Speko's own selections.
The timing is almost comically on-theme: the routing thesis arrives at the voice layer one day after Stripe reportedly agreed to buy OpenRouter for $7 billion plus. Every layer of the stack is growing a meter and a switchboard.
HN pushed back in two directions, and both are worth keeping. End-to-end speech models may kill the cascaded three-model architecture Speko optimizes (their counter: 95 percent of production systems are still cascaded), and local models may kill hosted routing entirely. Notice both objections are about whether the layer survives long-term, not whether it is valuable today. That was the OpenRouter objection too, and OpenRouter just sold for $7 billion.
Comments