AINEWS
Speko's pitch: voice AI is fragmented the way text models are, and the right product is no longer a model — it is the layer that picks the model on every call.
Voice AI has split into more than a dozen speech-to-text models, and the most accurate one in English is not the most accurate in Hindi. Cost and accuracy trade off differently in every language, and a Y Combinator S26 company called Speko is making the bet that, in that condition, the right product is no longer a model. It is the layer that picks the model on every call.
Speko's launch site publishes a public cost-versus-accuracy leaderboard, lists the company in YC's S26 directory, and ships a self-hostable open-source gateway on GitHub. The shape of the bet is the same one that produced OpenRouter for text-generation models. When the underlying engines fragment faster than any one of them can win, the durable product is the switchboard.
Per Speko's published benchmarks, Universal-3.5 Pro posts a 2.0% word error rate at $0.0075 per minute of audio, GPT-4o Transcribe lands at 2.3% WER and $0.0060 per minute, GPT-4o-mini Transcribe hits 2.7% and $0.0030, Qwen3-ASR sits at 2.8% and $0.0054, and Realtime STT-1 lands at 3.3% and $0.0025. The tail is where the dispersion shows. Nova-3 is at 9.8% WER and Ink-2 at 11.0%, both per Speko, and three other entrants the company lists carry no published WER figure at all, only per-minute price. A roughly 4x cost spread separates the cheapest engine from the most expensive in that table.
That is the structural condition Speko is naming, not a conclusion it has earned. The WER and dollar figures are Speko's own measurements under an undisclosed methodology, and the leaderboard is a marketing artifact as much as a dataset. The honest read is that variance exists, and the company is betting developers feel it.
The mechanism is small. Speko's gateway speaks the OpenAI API format, so an existing OpenAI speech-to-text, language-model, or text-to-speech client only needs a new base URL (https://api.speko.ai/v1) and a model: auto string in the request. On each call, the gateway selects the provider per stage, covering speech-to-text, language, and text-to-speech, based on the language, cost ceiling, and accuracy objective the developer passes in. There is no model-selection logic for the developer to maintain; the router owns it.
For teams that want to keep credentials and routing logic on their own infrastructure, the open-source Speko Gateway ships as a customer-side runtime. It exposes one local streaming protocol across voice providers, keeps bring-your-own-key credentials in-process, and optionally phones home to Speko for managed routing, observability, and consolidated billing. A public Docker image, spekoai/gateway:latest, is already on the registry, and the docs walk through dropping it into an existing LiveKit agent image with a few Dockerfile lines.
The marketing site also surfaces an MCP server at [mcp.speko.ai/mcp](https://mcp.speko.ai/mcp), installable as claude mcp add --transport http speko https://mcp.speko.ai/mcp. That is a small sign the company is targeting agentic-developer workflows, not just voice product teams.
A routing layer is a real product only if the routing decision is doing work a developer could not do by hand. The version Speko is launching does not yet make that case independently. The benchmark methodology behind the leaderboard is not visible on the public site. The early-preview status flag on the changelog explicitly warns that the protocol may change before the first stable release, which means anything wired to api.speko.ai/v1 today is a moving target. The "OpenRouter for voice AI" framing is Speko's own positioning, an analogy the company is inviting the market to use, not a category the market has already formed around.
There is also a thinner version of this story that does not need a startup at all. A large customer can write the same router in a weekend against OpenAI, Deepgram, Cartesia, and the other providers Speko lists on its models page. What Speko is selling is the maintenance: the benchmarks that stay current, the per-provider protocol drift handled in one place, the consolidated billing. That maintenance has to prove itself over months, not at launch.
The S26 placement and the public benchmarks put the company in front of developers who would otherwise build that router themselves. Whether the bet pays off depends on whether Speko's measurements stay ahead of the underlying providers' own releases, and whether the OpenAI-compatible base URL becomes a default that other voice stacks adopt, or stays a single-vendor convenience.