customers
All customer stories
Top Voice AI companies are building with Assembly.
resources
Latest Release
Text-to-Speech API
Our first text-to-speech model, AssemblyAI-TTS-1, is built for voice agents, on live calls. It sounds natural, responds fast, and keeps callers on the line.
resources
Speak to your app and have it speak back. Universal-3.6 Pro Realtime transcribes the user, LLM Gateway or your own model writes the reply, and the Text-to-Speech API voices it, all on one API key and one bill.
Hey, can you check where my order #48217 is?
Let me look—it shipped yesterday and arrives Friday.
Great, can you text the tracking link to my phone?
Universal-3.6 Pro Realtime hears the caller, your LLM decides what to say, and the Text-to-Speech API speaks the reply.
Built-in turn detection on the way in and about 90 ms time to first audio on the way out keep the back-and-forth feeling like a conversation.
The loop is built on Universal-3.6 Pro Realtime, so the words feeding your model are right the first time.
36%
improvement in close rate
“The new Universal-3.5 Pro speech model from AssemblyAI is best so far in terms of accuracy, latency, and language switching.”
80%
increase in customer satisfaction
“Assembly has saved us countless hours managing models, and provided exceptional accuracy.”
36%
improvement in close rate
“The new Universal-3.5 Pro speech model from AssemblyAI is best so far in terms of accuracy, latency, and language switching.”
80%
increase in customer satisfaction
“Assembly has saved us countless hours managing models, and provided exceptional accuracy.”
Start from the LiveKit or Pipecat reference implementation, which wires Universal-3.6 Pro Realtime, LLM Gateway, and the Text-to-Speech API together, so your first real-time exchange happens in minutes. Speech-to-text is $0.45/hr and Text-to-Speech is $15 per 1M characters, with voices in English, Spanish, French, German, Italian, and Portuguese. Try it in the playground with no code first.
Start building freeNo credit card required
A speech-to-speech loop is only as good as the words it hears. Streaming word error rate measures realtime Speech-to-Text accuracy at the front of the loop against a human reference—errors here compound through reasoning and the spoken reply.
Normalized WER on the voice-agent caller corpus, mean of three noise conditions.
*Lower is better*Source: AssemblyAI realtime benchmarks technical report v2026.09 (assemblyai.com/benchmarks), data 18 Aug–24 Sep 2026.
