Speko launches a benchmark-driven router that picks voice AI models per language

Speko, a Y Combinator S26 startup, is pitching itself as the OpenRouter for voice AI: a single gateway that sits in front of every major speech-to-text and voice model and routes each request to the best option. Its central claim is that model selection should be data-driven rather than marketing-driven — instead of trusting a vendor’s English-only leaderboard, Speko benchmarks each model language by language (English, Arabic, French, German, Hindi, and more) and chooses based on published word-error-rate and cost measurements per language and per objective. Its comparison table spans models from AssemblyAI, OpenAI, Alibaba’s Qwen, Google Chirp, Deepgram, ElevenLabs, xAI, and others, scoring accuracy against per-minute cost.

The product is built to be a drop-in replacement rather than a new SDK. Speko speaks the OpenAI API, so existing voice frameworks like LiveKit Agents only need a new base URL, an API key, and the model string set to “auto” to have STT, LLM, and TTS calls routed automatically. There is also an MCP endpoint for pointing agents directly at the service. The pitch targets teams building multilingual voice agents who currently have to test and wire up providers one by one.

The significance is less about a novel model and more about consolidation and transparency in a fragmenting speech-AI market. As the number of ASR and realtime-transcription providers grows, an abstraction layer that benchmarks them on a level, per-language basis and unifies billing and integration behind one key addresses a real operational pain point — though it also concentrates dependency on a single intermediary, and the value hinges entirely on how current and trustworthy its benchmarks stay.