Speko

congrats on the launch Bek! This is cool

Looks good, but would be nice if the site provided also actual voice samples of each model, e.g. just same text with type of voice prompt benchmarked accross different models.

This is super cool!

I love when a product does all the things I hate doing. Congrats, Speko. Proud user here.

Congratulations Bek. I have a question since Speko decides the stack before the session starts to keep audio paths direct, how do you handle mid-call failovers? If a specific provider suffers a sudden latency spike or an outage mid-session, does the system stay locked into that stack, or is there a way to redirect without dropping the user's call

What are the advantages compared to Google’s open-source libraries and other alternatives? Also, what about less commonly supported languages like Swedish or Danish?P.S. I just saw a post here about a kids’ app. You should consider partnering with them - they have voice recognition for children and convert speech to text.

Most of the audio I deal with is iPhone voice memos, never the clean format the docs assume. What happens to the input before it reaches a model?

Super hyped!

Congratulations on the launch @abdikbek! All of us needed a voice AI OpenRouter for sure. Wondering how often you would refresh this list of APIs or include more providers.

I’ve already discussed this direction with Bek, and we even hosted a deep-dive online meetup together about the development of voice AI.That’s why it’s especially interesting to see how the idea behind Speko is now evolving into a real product. Choosing the right voice stack is becoming increasingly complex as new models and providers keep emerging.I really like Speko’s approach: independent benchmarks, routing based on the specific use case and language, and no bias toward their own models.Very interesting product and direction. Good luck to Bek and the Speko team!

Hi Product Hunt! I'm Bek, founder of Speko.

Before this I was cofounded/CTO at a startup building voice AI apps. I left after Series A to work on the pain we lived with every week: we had to support 10+ languages, and I wished there was some kind of gateway that could route to the right voice tech based on the use case and language. Back then our process was a Google Sheet with 15 audio links we sent to people who speak those languages to score. If a new model was good, you swap. Every week something new comes out, even in English, and testing it is so much R&D cost that most teams end up running outdated models - swapping always looks like an R&D project. So early this year we started Speko.

What it is: a router for voice models. You come with your use case and language, and we route to the best speech-to-text, LLM, and text-to-speech for it - measured on our public benchmarks, not vendor marketing. You should be running the best voice stack at any given moment. We don't train or sell models ourselves, and that is precisely how we keep the rankings impartial.

How the approach evolved: we launched on Hacker News recently and the discussion sharpened how we explain this. "Router" hides three questions: what gets picked (a model, a provider, or the whole stack), where it lives (an external hop or inside your session), and when it decides (session start or mid-call). A voice agent is a live duplex session, not a batch request - that is why this is its own product. And we are proxy-less: routing is decided before the session starts, then audio flows directly to the provider - no extra hop in the audio path.