Speko

Teams building voice agents need to choose among dozens of STT, LLM and TTS models, but accuracy varies by language and no single provider wins across the board. The page shows that 10 of 53 benchmarked models are only measured in English, so their rank in other languages is unknown, including the top English model. It also shows 4 different models win across 17 languages, meaning the right pick changes depending on the language a product's users speak. Without a way to compare models on accuracy (WER), latency and cost per stage, teams building multilingual voice agents are left guessing or locked into one provider.

Speko offers a hosted, provider-neutral data plane (the Router) at router.speko.dev that wires speech-to-text, LLM and text-to-speech models into one API with typed contracts and managed routing. It publishes benchmark coverage by language and model, including word error rate and finalize latency (p50), so developers can score models against cost per stage (STT, LLM, TTS, and speech-to-speech). A separate open customer-side runtime, the Gateway, supports LiveKit and Pipecat with provider-direct streaming, local BYOK credentials, and optional Speko-managed routes, with native integrations shown in Python code samples. The Router is also exposed via public OpenAPI and AsyncAPI contracts and an MCP endpoint that can be added directly to tools like Claude.