Speko

Speko is a router for voice AI models. It benchmarks speech and language models language by language and routes each voice session to the model that performs best for the language being spoken, behind a single API. The vendor's premise is that most voice models are benchmarked only in English, so the best choice changes with the language a product's users speak. The hosted router sits between an application and its providers for speech to text, language model and text to speech stages. Speko publishes benchmark tables covering dozens of models from providers such as Deepgram, OpenAI, ElevenLabs, AssemblyAI and Google across 17 languages, with accuracy measured as word error rate and latency measured as time to finalize. The service offers measured model selection and failover before a response begins. Developers can get an API key from the site, and the vendor also offers an onboarding path aimed at coding agents. Pricing has three paths. The Router option adds 5 percent on top of the provider's published rate. The Speko infrastructure option bundles speech to text, language model and text to speech for $0.09 per minute. Enterprise pricing is custom, with a committed monthly minimum, dedicated support and a contractual service level. Every account receives $10 in signup credit. Access is currently a public preview, and the preview carries no service level agreement. The company states it is backed by Y Combinator.