The archive · AI & Models · Technical decision · 2026
Speko bets developers will pay for benchmarked voice-AI routing
YC S26 startup benchmarks STT, LLM, TTS models language-by-language and routes calls to the winner via one API, after four years of manual bake-offs at Hupo.
Speko
What the business is
A hosted router for voice AI: one API key routes STT, LLM, and TTS calls across vendors based on public, language-specific benchmarks (word error rate, latency, cost per minute), plus an open-source gateway for LiveKit and Pipecat.
Starting capital：$1.1M pre-YC from US funds and angels plus a $500,000 YC standard commitment, roughly $1.6M total.
How it started
Beknazar Abdikamalov founded Speko in 2026 in San Francisco after four years as co-founder and CTO of Hupo, where he manually benchmarked STT, LLM, and TTS providers for production voice and found the best combination changed monthly, especially outside English.
What happened
Speko joined Y Combinator's Summer 2026 batch with a four-person team, raised $1.1M from US funds and angels before YC plus the $500K standard commitment, and launched with public benchmarks across nine languages and more than 20 STT models, an MIT-licensed open-source gateway, and LiveKit and Pipecat integrations.
How it ended up
Still live: the hosted router runs at speko.ai with public leaderboards at benchmarks.speko.ai, and no revenue or customer figures have been disclosed.
Background
Speko is a YC Summer 2026 startup building a routing layer for voice AI. A voice agent chains three model stages - speech-to-text, an LLM, and text-to-speech - and each stage has a dozen credible vendors whose rankings change every month. Speko benchmarks them language by language, publishes the measurements, and routes a customer's calls to the current winner through one API key.
Founder Beknazar Abdikamalov built the company out of a problem he had as CTO of Hupo, where he manually ran voice-model bake-offs for enterprise clients like Morgan Stanley, Prudential, HSBC, and Grab. He found that most teams evaluate once, ship a stack, and never recheck, so production agents quietly run on last quarter's models.
Speko raised about $1.1M from US funds and angels before joining YC, received the standard $500K YC commitment, and launched with public leaderboards covering nine languages and more than 20 speech models, an MIT-licensed open-source gateway, and LiveKit and Pipecat integrations.
Its Launch HN on August 17, 2026 hit the front page with 118 points and 69 comments. Abdikamalov reported external usage growing about 25% per week since the batch started, and the company is betting developers will pay to outsource the evaluation and routing that almost nobody re-runs.
What has to be true
- Voice stacks rot silently: teams pick a model combination once, and every new release makes the choice more outdated without anyone noticing.
- Benchmark credibility is the product: publishing WER, latency, and cost per minute language by language turns measurement itself into a reason to visit and a reason to trust routing.
- The aggregator model is proven: OpenRouter showed developers will standardize on one key across providers, and Speko applies the same pattern to the three-stage voice pipeline.
- A founder with four years of manual bake-offs means the company is selling a solved personal problem, not a hypothetical pain point.
- Open-sourcing the gateway removes the adoption risk of a new network hop and keeps BYOK users in control of credentials.
What can be applied
A founder problem plus a defensible measurement surface beats a generic AI wrapper: Speko sells the benchmark itself as the product and makes trust in its numbers the moat.
Aftermath
As of September 2026, Speko is an active YC S26 company with a four-person team in San Francisco. The hosted router is live with public benchmark boards, and the open-source gateway supports LiveKit and Pipecat. It has not disclosed revenue or paying-customer numbers, and its main challenge is proving its measurements stay accurate enough to justify charging for the traffic it routes.
Sources
- Launch HN: Speko (YC S26) - OpenRouter for Voice AI
- Speko routes voice AI across providers using language-specific benchmarks
- Speko: OpenRouter for Voice AI - one API, every call on the best speech model
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card