Today, we’re launching the Streaming API: stream speech as it’s generated, transcribe audio in real time, and control how voices speak with Voice Instructions now in beta.

👉 Try the API Playground or read the API documentation.

Voice AI That Keeps Up With the Conversation

Voice AI is only as useful as the time it takes to respond.

With traditional text-to-speech, you send a complete sentence or paragraph, wait for the audio to finish generating, and only then can playback begin. For a short message, that delay might be acceptable. For a voice assistant, phone system, or live conversation, it changes the experience entirely.

Streaming removes that wait.

With the Streaming API, audio can start playing while the rest is still being generated. At the same time, Realtime ASR can turn speech into text while someone is still talking.

That means developers can now build genuinely interactive Khmer voice experiences voice agents, live captions, voice interfaces, and IVR systems without stitching together multiple tools that weren’t designed with Khmer in mind.

Streaming TTS Start Speaking Before Generation Finishes

KiriTTS now supports streaming.

Send text to the API and audio chunks are returned as they’re generated. Your application can begin playback immediately instead of waiting for the entire audio file to be produced.

It works through the same OpenAI-compatible endpoint you already know:

curl -X POST https://api.kiritts.com/v1/audio/speech \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kiritts",
    "input": "សួស្តី! ខ្ញុំឈ្មោះ Maly។",
    "voice": "Maly",
    "stream": true
  }'

That’s it. One flag turns on streaming.

If you’re already using the Kiri API, you can add streaming without changing your integration. And because the API is OpenAI-compatible, existing OpenAI SDK-based applications can point to Kiri’s API endpoint and use the same familiar interface.

Where streaming TTS shines

Voice assistants respond at the pace of a natural conversation
Phone and IVR systems reduce silence between a caller’s request and the system’s response
Interactive applications start speaking without waiting for a complete response
Long-form narration begin playback immediately while the remaining audio is generated

Realtime ASR Understand Speech While It’s Happening

Voice conversations work both ways.

That’s why we’re also introducing Realtime ASR: persistent, low-latency speech recognition that turns incoming audio into text as the person speaks.

Instead of uploading a completed recording and waiting for a transcription, your application sends audio as it’s captured. Kiri returns partial transcription results as speech is recognized, refining them as the sentence continues.

The result is an experience that feels much closer to a live conversation.

Where Realtime ASR shines

Live captions caption meetings, streams, classes, and events in Khmer
Voice typing turn natural Khmer speech into text as you speak
Voice agents begin processing a request before the speaker has finished
Interactive applications react to speech in real time rather than after a recording ends

Batch transcription isn’t going away. It remains the right choice for recorded audio and completed files.

Realtime ASR is for when the audio is happening now.

Voice Instructions (Beta) Tell the Voice How to Speak

A voice shouldn’t just say the right words. It should say them the right way.

Voice Instructions lets you describe how you want a voice to speak using plain language.

For example:

“Speak softly and slowly, like telling a bedtime story.”

Or:

“Sound like an excited sports commentator.”

No SSML. No complicated parameter matrices. Just describe the delivery you want, and Kiri uses that instruction to shape the performance.

Voice Instructions can be applied per request through the API and per paragraph in the Kiri TTS Studio. That means a single script can move naturally between different speaking styles from a calm narrator to an energetic announcer.

Voice Instructions is currently in beta. It’s live and ready to experiment with, and we’ll continue improving it based on what developers create and where the model still falls short.

We’d love to hear what you build and what you want it to do better.

Built for Developers

The Streaming API is part of the developer platform we’ve been building to make voice AI easier to integrate into real applications.

OpenAI-compatible APIs

Use familiar SDKs and patterns without learning an entirely new API from scratch.

Projects and API keys

Create keys for different applications, rotate credentials, and keep track of usage and logs at the project and key level.

Live API Playground

Test streaming TTS and Realtime ASR directly in your browser before writing a line of code.

Explore the API Playground.

Get Started in Three Steps

1. Create your account

2. Create an API key

Open the API dashboard and create your first key.

3. Make your first streaming call

Follow the API documentation for copy-and-paste examples in cURL, Python, and Node.js.

Khmer Voice AI, in Real Time

Real-time voice AI shouldn’t be limited to a handful of global languages.

With the Streaming API, Khmer developers now have the building blocks to create voice experiences that respond as conversations happen:

Speech that starts immediately. Transcription that keeps up. Voices that take direction.

From voice agents and customer service systems to live captions, education, media, and entirely new applications we haven’t imagined yet we’re excited to see what you build.

Start building with the Streaming API →