Top AI Product

Stripe just paid $7B+ for OpenRouter. Speko’s bet: the same routing-layer play, but for voice. Its Launch HN hit the front page this week.

One API for 15+ voice providers

Speko is an API gateway that routes each call across 15+ STT, TTS, and voice-to-voice providers — ElevenLabs, OpenAI, Cartesia, Deepgram, Rime, Hume — picking the best model for your language, latency, cost, or accuracy target. Founder Beknazar Abdikamalov spent four years building voice agents across Asia in 10+ languages, where model quality swings wildly and nobody re-benchmarks every quarter.

The API is the product

One unified endpoint plus a TypeScript SDK: declare a constraint (accuracy, latency, cost, or balanced), Speko returns the winning provider in response headers, with server-side failover. Session plans are prefetched, so routing adds zero round trips mid-call. The gateway is MIT open source; the hosted router is the paid tier. One medical team found STT word error rates ranged 8–19% across providers.

Why the data is the moat

Continuous benchmarks on 61 voice models across 10 languages, published openly — including cases where their router picked wrong. Top models hit 94% turn-taking accuracy vs 46.9% for basic VAD timers. Usage is up ~25% per week since late June. HN skeptics say end-to-end speech models will kill the cascade; enterprise operators counter that 95% of production voice is still cascaded. Either way, someone has to keep score.

You Might Also Like

• Grok stt 1 0 Hits Openrouter Xais Transcription api Undercuts Deepgram at 0 10 Hour
• Assemblyais Voice Agent api Undercuts Vapi and Retell With 307ms stt Latency
• Openai gpt Realtime 2 Translate Whisper Three Voice Models one api Several Startups Erased
• Kani tts 2 Just Dropped and it Only Needs 3gb of Vram to Clone Your Voice
• Hume ai Open Sources Tada an llm Based tts With Zero Hallucinations and 0 09 rtf