Building a voice agent that works across languages forces a hard question early: which text-to-speech engine do you build on? The demos all sound great. Production is where they separate — under real load, across accents, at a cost you can sustain. Here is what we learned comparing the field.
What we were actually optimising for
- Voice quality — natural prosody that holds up across languages, not just English.
- Latency under load — time-to-first-audio when many calls hit at once, not in a quiet benchmark.
- Cost at scale — per-minute economics that survive real call volume.
- Multilingual coverage — the specific languages and accents our users actually speak.
The trade-off map
| Priority | What gives | What to watch |
|---|---|---|
| Best voice quality | Often higher latency / cost | Streaming and caching to hide it |
| Lowest latency | Sometimes thinner voices | Quality floor on key languages |
| Lowest cost | Coverage and naturalness vary | Test your languages, not the demo’s |
The lessons
- Benchmark under concurrency — idle latency tells you nothing about a busy line.
- Stream audio so perceived latency stays low even when generation isn’t instant.
- Test on your real languages and accents; aggregate scores hide the gaps that matter.
“A TTS engine that sounds perfect in a demo can fall apart on a busy multilingual line. Choose for the load you’ll actually carry.”
KnackLabs

Kartik Bansal is the Founder and CEO of KnackLabs, an AI transformation firm that embeds senior engineers inside businesses to build and run production AI. Before KnackLabs he was a senior engineer at Microsoft and co-founder and CTO of Convosight.



