A New Startup Wants to Pick the Best Voice AI for You Automatically
On August 17, 2026, a startup called Speko launched publicly through a post on Y Combinator's Hacker News. Speko's idea is straightforward: when you build a voice assistant, you need three pieces — software to turn spoken words into text, an AI to generate a response, and software to turn that text back into speech. Lots of companies make each piece, and which one is best depends on what you care about: accuracy, speed, cost, or a balance. Speko picks the right combination for each request, and you access all of them through a single API key.
Speko's founder, Bek, previously spent four years as cofounder and CTO building enterprise voice agents across Asia in more than ten languages.
Here is how it works. You tell Speko what you want optimized — accuracy, latency, cost, or a balance — along with the language and region. Speko tests the available models, picks the best one for your criteria, and sends back the response along with details about which provider and model it used and how they scored. The company publishes those test results openly on its website, running the same inputs across all models in one region on dated runs. Notably, the published results include cases where Speko's own picks performed worse than other options. The tests cover spontaneous speech, money and dates, and ten-minute real-world recordings, and model rankings shift depending on the test.
For text-to-speech quality, Speko trained an automated scorer using blind listening votes — people compared two voices and picked the better one without knowing which provider made each. That automated scorer agrees with human raters about as often as raters agree with each other, even for providers it has never been tested on. Speko does not make or sell its own AI models; its entire value is in choosing, routing, and testing the ones others have built.
Speko open-sourced its gateway software at github.com/SpekoAI/gateway under the MIT license, meaning anyone can use and inspect it freely. The gateway is a small program that runs alongside your application on the same machine, talking to it over a fast local channel. It holds your keys, connects directly to whichever provider Speko selected, and skips a step that would otherwise slow things down at the start of a conversation. If the chosen provider refuses the connection, the gateway falls back to the next-best option — but only at the moment a session starts, not midway through.
In BYOK mode — bring-your-own-key, meaning you use your own accounts with each provider — the gateway never talks to Speko's servers at all. Anonymous, content-free telemetry is on by default and can be turned off with a single setting. 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, calling Speko "OpenRouter for Voice" on Y Combinator. OpenRouter is a known service that routes text-based AI requests across providers; Speko aims to do the same for voice. 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 these architecture choices matter. Voice assistants are far more sensitive to delays than text chatbots. A half-second pause feels natural in text but awkward in conversation, so Speko designed its gateway to run locally and skip unnecessary network steps at the start of each session. Limiting failover to the connection-setup stage is a practical tradeoff: it avoids the messiness of swapping providers mid-utterance, at the cost of not recovering from failures that happen after a session is already live. For real-time voice, that is a reasonable line to draw.
The BYOK air-gap is the other decision worth noting. Companies evaluating voice AI often worry about sensitive audio passing through a third party's servers. A gateway that never contacts Speko in BYOK mode removes that concern entirely, and because the code is open source, the claim can be verified. The telemetry default — anonymous, with no content, turned off with one setting — is a reasonable middle ground, though security-conscious teams will appreciate that the BYOK path involves genuinely zero contact with Speko.
Publishing benchmark results that include cases where Speko's own picks lose is an unusual choice. Companies that sell model selection have a natural incentive to show only results that make their picks look good. Showing the losses builds trust with the people Speko needs most: engineers who will check the methodology before they rely on the routing decisions.
The "OpenRouter for Voice" label fits in one way and falls short in another. OpenRouter routes across providers for text generation, where speed and coordination are simpler. Voice pipelines chain three different types of AI models, each with different ways of failing, different speed characteristics, and different quality tradeoffs. Routing across that chain is a harder problem, and one that shifts every time any provider releases a new model.
What we do not yet know is whether the hosted service can keep growing as real production workloads put pressure on the testing and selection systems. Early growth numbers from a YC batch are encouraging but not predictive. The open-source gateway, the public benchmarks, and the zero-contact BYOK mode together suggest Speko is betting that transparency and the ability to self-host will earn enterprise trust faster than a closed product. That bet aligns with how infrastructure software has typically gained adoption in this industry.
