Async Voice AI

Speed is the whole point here. Async says its model starts audio faster than the low-latency leaders, with a median time to first byte around 166 ms and total latency under 300 ms. For a voice agent that means the reply lands before the silence gets uncomfortable. The company measures this itself, so treat the exact numbers as its own benchmarks, but the streaming architecture backs up the claim.

You can clone a voice from a 3-second sample. No training run, no waiting queue, no separate fee on the developer plan. Upload a clip and you get a voice that keeps the tone, accent and pacing of the original. It's the feature that turns a generic TTS API into something you can brand, whether that's a character, a host or a company voice.

Async splits its text-to-speech into two lines. Async Pro v1.0 is the quality pick and handles English best, with clean treatment of dates, numbers and currencies. Async Flash v1.5 trades a little polish for speed and pairs well with live agents. There's also an older Flash v1.0 with the widest language coverage. Picking between them is really a choice between best sound and fastest reply.

The API covers more than 15 languages with what the company calls native pronunciation, and the same endpoint serves them all. Text normalization handles the messy parts, like reading dates, currencies and abbreviations the way a person would. For teams shipping in several markets, that means one integration instead of a different model per region.

Text-to-speech comes in four shapes: a plain HTTP call that returns one audio file, a streaming endpoint that starts playback before generation finishes, a timestamped version for captions, and a WebSocket for feeding text in as an LLM writes it. Speech-to-text works the same way, with a file upload for recordings and a live stream for real-time transcription. Few other APIs give you this much choice over how audio reaches your app.

Beyond the core models, Async exposes the small knobs that decide whether speech sounds produced or natural. You can set custom pronunciations with IPA phonemes for brand names and acronyms, read numbers digit by digit for phone numbers, and insert exact pauses with a break tag. For wiring it in, there are ready integrations for Pipecat, LiveKit, Twilio, n8n and Picsart Flow, so you're not writing the plumbing from scratch.