Speko launches a benchmark-based router for voice AI models
The YC-backed gateway picks speech providers by language, quality, latency and cost, then handles failover through one API.
By Ryan Merket · Published
Why it matters
Voice AI stacks can age quickly as model quality and pricing change. Speko wants to make continuous evaluation and provider switching an infrastructure function.
Beknazar "Bek" Abdikamalov (@beknabdik) launched Speko on July 29th, giving voice AI developers a single gateway that selects speech and language models using continuously updated benchmarks.
https://x.com/beknabdik/status/2082544456604287230?s=46
Abdikamalov described Speko as "OpenRouter for Voice" in the launch announcement. His premise is that speech-to-text and text-to-speech models arrive and improve too quickly for developers to keep rerunning vendor comparisons. Speko measures providers by language and routes requests according to the results.
The San Francisco operation is part of Y Combinator's Summer 2026 batch, where Speko is listed with four people. Abdikamalov previously co-founded Hupo and served as its chief technology officer after working as a software engineer at Amazon. He holds bachelor's and master's degrees in computer science from Ulsan National Institute of Science and Technology in South Korea, according to YC.
Speko is Abdikamalov's attempt to own the layer between voice applications and a growing collection of model providers. Voice agents typically combine speech recognition, a language model and speech synthesis. Each stage introduces separate decisions around accuracy, response time, language coverage, price and reliability. Changing one provider can require new credentials, integration work and testing.
How the router works
Speko exposes speech-to-text, language-model and text-to-speech endpoints through its TypeScript SDK. A developer sends a routing intent containing a language and, optionally, a region and an optimization target. Speko scores eligible providers, selects the highest-ranked option and moves to the next candidate if the first provider fails.
Responses include the provider, model, failover count and benchmark run used for the decision. Developers can also restrict the eligible providers when they need to meet compliance rules, cap costs or pin a model during debugging.
That makes Speko a routing and evaluation layer rather than a closed voice-agent stack. Speko can plug into frameworks including LiveKit and Pipecat, while its own platform also supports browser conversations and inbound and outbound phone agents.
The public model directory lists 53 entries: 13 speech-to-text models, 16 language models, 15 text-to-speech models and nine speech-to-speech models. Speko separately claims coverage of more than 50 providers and more than 140 models across at least 10 languages, indicating that its marketed catalog is broader than the set displayed in the directory.
Speko's benchmark boards publish measurements including word error rate, finalization latency, time to first token and estimated cost. The rankings vary materially by metric. A model that performs well on transcription accuracy can trail another provider on streaming latency, leaving the router to make a workload-specific choice instead of declaring one universal winner.
The business model
Speko is entering public preview with three self-service price points. Its pricing page offers $100 in signup credit, a router plan charging 5% above the underlying provider rate and a Speko-hosted infrastructure plan priced at $0.09 per minute. None of the public-preview plans includes a service-level agreement. Enterprise contracts add custom pricing, dedicated support and negotiated service terms.
Y Combinator is Speko's publicly identified backer. YC lists Speko in its Summer 2026 batch; YC's standard deal invests $500,000, split between $125,000 for 7% and $375,000 through an uncapped most-favored-nation SAFE.
The routing pitch places Speko alongside voice-agent infrastructure providers such as Vapi, Retell AI, LiveKit Agents and Pipecat. Those products span orchestration, telephony and application frameworks. Abdikamalov is making a narrower infrastructure bet: developers will want model choice and independent measurements without maintaining a separate integration for every speech vendor.
That bet depends on the benchmark layer remaining credible as the catalog expands. The router is useful only when its tests resemble production traffic across accents, languages, network conditions and conversational workloads. Speko's decision to publish dated results and expose the benchmark run behind each route gives developers a way to inspect those choices. It also turns the benchmark methodology into the core product rather than a marketing attachment.