Speko Launches a Routing Layer for Voice AI Pipelines
Speko, a Y Combinator S26 startup, launched publicly on August 17, 2026 via a Launch HN post. The company has built a routing layer for voice AI pipelines — software that automatically picks the best combination of speech-to-text, large language model, and text-to-speech services for each incoming request, all accessible through a single API key. Its founder, Bek, previously spent four years as cofounder and CTO building enterprise voice agents across Asia in more than ten languages.
The core product is a router API. You send it optimization criteria — accuracy, latency, cost, or a balance — along with language and region, and it filters and selects models accordingly. Speko benchmarks candidate models, picks a winner, and returns responses with HTTP headers identifying the provider, model names, and scores. The company publishes those benchmark results publicly, running identical inputs across all models in one region in dated runs. Notably, the published boards include cases where Speko's own selections perform worse than alternatives. The benchmarks cover spontaneous speech, money and dates, and ten-minute production takes, and model rankings shift based on those tests.
For text-to-speech quality, Speko trained an automated scorer for TTS naturalness using blind head-to-head listening votes. That scorer picks the same winner as human raters about as often as raters agree with each other, including on providers it has never seen votes for. Speko does not train or sell models itself; the entire value proposition sits in selection, routing, and benchmarking.
Speko open-sourced its gateway as a Go binary at github.com/SpekoAI/gateway under the MIT license. Think of the gateway as a local intermediary that runs alongside your application inside the same container. It speaks one local protocol over a Unix socket — a fast communication channel between processes on the same machine — pins provider hosts, and attaches the user's keys. It prefetches signed session plans so a new session connects to a provider directly from memory, avoiding a control-plane round trip while a caller waits. Failover operates only during connection setup: if the chosen provider refuses the connection, the gateway falls back to runner-up providers.
In BYOK (bring-your-own-key) mode, the gateway does not communicate with Speko's cloud at all. Anonymous, content-free telemetry is enabled by default and can be disabled with a single environment variable. The gateway and BYOK setup are free indefinitely; Speko charges for its hosted router and managed keys, with consolidated billing across providers.
Bek announced the product at speko.ai on his X account on July 29, 2026, positioning Speko as "OpenRouter for Voice" on Y Combinator. YC's own company profile describes it as a router for voice models, and the accelerator published a LinkedIn post noting one API key for speech-to-text, LLM, and text-to-speech access. Since starting the YC batch in late June 2026, Speko's external usage has grown roughly 25 percent week over week.
The broader context here is that the architecture decisions deserve attention. Running the gateway as a sidecar over a Unix socket means latency-sensitive voice loops get local inter-process communication rather than a network hop to a hosted router. Prefetching signed session plans eliminates the control-plane round trip at session start, which is the right place to optimize — voice agents are unforgiving of cold-start latency in ways that text chat is not. And constraining failover to the connection-setup stage, rather than mid-session, is a pragmatic tradeoff: it avoids the complexity and inconsistency of swapping providers mid-utterance, at the cost of not recovering from provider failures that occur after a session is live. For real-time voice, that is a defensible boundary.
The BYOK air-gap is the other decision that stands out. Enterprises evaluating voice AI are frequently concerned about sensitive audio flowing through an intermediary's cloud. A gateway that never contacts Speko's servers in BYOK mode removes that concern entirely, and the MIT-licensed binary means the claim is auditable. The telemetry default — anonymous, content-free, opt-out via one environment variable — is a reasonable middle ground for a hosted product, though security-conscious teams will appreciate that the BYOK path is genuinely zero-contact.
The public benchmark boards, including cases where Speko's picks lose, are an unusual transparency choice. Companies that sell model selection have a structural incentive to publish only results that validate their selections. Showing the losses builds credibility with exactly the audience Speko needs — engineers who will check the methodology before trusting the routing decisions.
The "OpenRouter for Voice" framing is apt in one important sense and imperfect in another. OpenRouter routes across LLM providers for text generation, where latency and multimodal orchestration are lower-stakes concerns. Voice pipelines chain three distinct model categories — speech-to-text, LLM, and text-to-speech — each with different failure modes, latency characteristics, and quality dimensions. Routing across that chain is a harder problem, and one that changes with every model release from every provider in the stack.
What remains unclear is whether the hosted router can sustain its growth trajectory as production voice workloads put real load on the benchmarking and selection infrastructure. Early-stage growth numbers from a YC batch are encouraging but not predictive. The open-source gateway, the public benchmarks, and the BYOK air-gap together suggest Speko is betting that transparency and self-hostability will win enterprise trust faster than a closed hosted product. That bet aligns with how infrastructure software has typically gained adoption in this industry.
