Beknazar Abdikamalov founded Speko in 2026 as a San Francisco-based voice AI routing layer that selects speech, language and synthesis models using language-specific benchmarks. Speko provides a /v1
API, with one Speko key and benchmark-based routing across providers by language, latency, accuracy and cost.
Speko grew out of a problem Abdikamalov encountered before starting Speko, which joined Y Combinator's Summer 2026 batch. YC's current company profile lists a team size of four. Abdikamalov spent years building voice features while serving as co-founder and chief technology officer of Hupo, the Singapore leadership coaching platform that began as an employee mental-health product called Ami. At Hupo, Abdikamalov had difficulty determining which speech-recognition and synthesis providers worked best for particular languages, according to a Digital Business interview.
Abdikamalov has said he has worked with voice technology since 2019, according to a June 2026 interview with Digital Business.
That experience shaped Speko's product thesis. Voice agents typically chain at least three technical stages: speech-to-text, a language model and text-to-speech. Each stage can come from a different vendor, and the best-performing combination can change as providers release new models or adjust their infrastructure. Abdikamalov is betting developers will pay Speko to keep running those comparisons and update its routing automatically.
A founder problem turned into infrastructure
Abdikamalov grew up in Nukus, Uzbekistan, began programming in sixth grade and later moved to South Korea, according to a published interview. He earned bachelor's and master's degrees in computer science from UNIST after studying mathematics and computer science as an undergraduate. His career included engineering roles at Korean startups, including Tridge, and a software engineering job at Amazon.
He left Amazon to build Ami with former colleague Justin Kim in 2022. Ami later became Hupo, which focused on leadership development and management coaching. Hupo raised about $4 million from backers that YC identifies as including DST Global, Meta, Goodwater Capital and Collaborative Fund.
The path from Hupo to Speko matters because Abdikamalov was already buying and testing the services Speko now wants to route. In a June interview with Digital Business, he said the weaknesses were most visible outside English, where nominal language support could obscure lower accuracy or reliability in production.
Speko's public benchmark page is designed to make that gap visible. The homepage currently shows nine languages and a selector containing 21 speech-to-text models. Speko publishes measurements including word error rate, finalize latency, time to first token and cost per minute.
The displayed English table, for example, assigns AssemblyAI's Universal-3.5 Pro a 2.0% batch word error rate at $0.0075 per minute. OpenAI's GPT-4o-mini Transcribe is listed at 2.7% and $0.0030 per minute. The difference illustrates the decision Speko wants to automate: the most accurate option on a given test may carry a materially higher usage cost, while the cheapest model may require an accuracy tradeoff.
Those figures are Speko's measurements. They should be treated as company-published rankings rather than independent validation. Speko's commercial position depends on developers accepting its tests as a better basis for routing than the vendors' own leaderboards, which makes benchmark credibility central to the product.
One Speko key across providers
Speko borrows the aggregator model that OpenRouter, a gateway for language models, uses for provider selection, then applies it across the voice pipeline. OpenRouter can order providers by price, throughput or latency and fall back when an endpoint fails. Speko adds speech-specific measurements and language selection across transcription and synthesis.
Speko's documentation says developers authenticate with a Speko API key and call Speko's own /v1
endpoints rather than integrating separately with each upstream service.
For LiveKit, Speko documents a real-time speech-to-text, language-model and text-to-speech pipeline using its adapter on a LiveKit Agents worker. The integration lets developers configure Speko within an existing voice-agent project while Speko handles provider routing behind the API.
Benchmarking is also Speko's customer acquisition engine
The public rankings do two jobs. They give voice developers a reason to visit Speko before choosing a provider, and they produce the measurements Speko uses to justify its paid routing decisions. Each new model release creates another comparison for Speko to publish and another opportunity to persuade developers that maintaining an internal evaluation suite is wasted engineering time.
YC gives Speko capital and a concentrated customer pool
Digital Business reported that Speko raised $1.1 million from US funds and angel investors before YC. YC's standard deal commits another $500,000 through two safes, bringing the reported amount raised or committed to roughly $1.6 million. Speko has not identified the pre-YC investors in the sources reviewed, and customer, revenue and usage figures are unavailable.
YC also supplies a concentrated pool of potential buyers: startups building phone agents, customer-support tools and other voice products frequently face the provider-selection problem Speko is trying to remove.
Speko's immediate task is turning a useful research surface into infrastructure developers trust inside live conversations. Abdikamalov has experience buying and testing the services Speko now routes. The harder work is proving that Speko can measure changes in provider performance accurately, switch providers without degrading calls and remain credible while charging for the traffic it directs.