Speko says it solves a very real operational problem in voice AI: production systems are usually stitched together from speech-to-text, an LLM, and text-to-speech, and teams rarely revisit those choices because every model swap turns into another integration project. The product combines public benchmarks with a router that picks a stack based on constraints like accuracy, latency, cost, language, and region. It also ships an open source sidecar gateway for teams that want provider switching without putting another network hop in the audio path or handing over API keys.
The strongest reaction was that the benchmark layer is the compelling part. People wanted to know how Speko measures things like turn-taking, naturalness, and transcription quality. The founder’s answers landed because they were concrete. Turn-taking is scored mechanically on 200-plus real clips by judging whether the model correctly detects a finished utterance versus a pause.
TTS naturalness still uses blind listening rounds with native speakers. Domain vocabulary came up fast, and that exposed why generic leaderboard numbers are not enough. A model that wins on clean speech can lose badly on medical or product-specific terms, which makes continuous testing on your own data more valuable than a single headline score.
A second thread pushed on the architecture itself. Some argued the market is moving toward end-to-end
speech-to-speech systems, so routing separate
STT, LLM, and TTS vendors may be a temporary layer. The more grounded view was that enterprise buyers still overwhelmingly want cascaded systems because they are easier to control, observe, tune, and swap piece by piece. Even the founder’s own benchmark results reinforced that. Closed speech-to-speech models are improving, but long multi-turn calls still seem to break the open ones more often, and most production deployments today still prefer the composable stack.
The other persistent tension was local versus hosted. A few commenters dismissed paid voice vendors as a dead end and argued that on-device or
open-weight models will win. That did not really overturn the launch thesis. It sharpened it. Single-user dictation and translation may move local faster, but production phone agents still have tighter latency budgets, multiple model stages, and big quality swings by language and use case. Speko’s position only holds if it keeps measuring both hosted and open models in a way buyers trust. That makes this look less like a pure router and more like a benchmarking business with routing attached so the benchmark can actually change what runs in production.
speko.ai