Does speech to speech actually make voice agents better or just faster?

I’m Somil, I work on voice agents at Vaani in Bengaluru.

Something I read this week bugged me. One benchmark across 28 voice agents found the median one did no better than a plain words only pipeline. Meanwhile every new duplex model promises it hears emotion, sarcasm and hesitation that a transcript throws away.

Both can be true, but it made me wonder how much of what we call “feels more human” is just lower latency and better interruptions, and how much is the model understanding the caller better.

From my side, speech to speech models still seem less smart than the text LLMs they’re built on. For anything involving tools, account lookups or a long conversation, I’d still take the smarter model over the more natural one.

So for people who’ve tried both:

• Did you notice a real quality difference, or does it just feel faster?

• Where did the text pipeline beat s2s, or the other way round?

• If you could only fix one, would it be intelligence or turn-taking?