Today, we are excited to introduce Kiri-S2S (Kiri) as a Research Preview! It is the first Real-Time Speech-to-Speech (S2S) AI Model for Khmer, allowing you to speak naturally in Khmer and receive instant spoken responses.

👉 Try the live demo now: https://s2s.blizzer.tech/ — no sign-up, just talk.

Why Does Khmer Need Its Own AI?

For millions of people around the world, talking to AI has become second nature. You ask a question, and it responds in a matter of seconds.

But for the 17 million Khmer speakers, that experience has always felt like using a borrowed system built for someone else.

Khmer is too often overlooked. Most voice assistants don't support Khmer at all, or they speak slowly and roboticallyas if every word has to be translated to English behind the scenes before the system agrees to answer.

We believed that wasn't good enough.

Starting today, you can speak Khmer directly to AI, and it will respond in Khmer instantly, naturally, and in real time.

Meet Kiri-S2S (Kiri) the first native Speech-to-Speech (S2S) model for Khmer. Audio goes in, and audio comes directly out, with no intermediate text conversion steps in between.

Kiri-S2S Features & Capabilities

Natural Real-Time Conversation: No push-to-talk required. It responds in less than 1 second, and if you interrupt mid-sentence, it stops speaking and listens to you immediately.
Listens, Thinks, and Speaks in One System: It doesn't need to write down your words or read off a script. It understands tone, inflection, and emotion (such as hesitation, excitement, or curiosity) directly and responds like a real human.
Bilingual (Khmer ↔ English): Seamlessly switches between Khmer and English whenever you switch languages, and can even translate live during conversations.
4 Distinct Voices: Choose between Bosba, Borey, Setha, and Maly.
Supports Both Voice and Text: If you prefer typing, you can send and receive text messages just as easily.

The Moment We Knew This Mattered

There is something truly remarkable about hearing an AI respond to you in your native tongue.

Not because it translated a sentence from English. Not because it’s reading from a script. But because it truly understood what you said and chose how to respond on its own.

When Kiri-S2S pauses for a split second before answering a tough question or adjusts its tone naturally, it doesn't feel like three separate computer programs glued together.

It feels like a single brain.

How Traditional Voice Assistants Work (The Relay Race)

Historically, talking to a voice AI assistant felt like watching a relay race between three runners:

Runner 1 (Speech-to-Text): Converts your spoken voice into written text.
Runner 2 (AI Brain): Reads that text and drafts a written answer.
Runner 3 (Text-to-Speech): Reads that written answer back to you out loud.

Every time the baton is passed, valuable time is lost. Worse yet, the human identity of your voice is lost at Step 1including your tone, hesitations, and emotion.

For Khmer, if any single runner makes a small mistake, the entire response falls apart.

The Kiri-S2S Approach: Listen, Think, and Speak as One

Kiri-S2S eliminates the line of middlemen entirely, replacing them with a single unified intelligent system.

It is a native Speech-to-Speech Model that takes raw audio input, processes the meaning, and generates audio output — simultaneously, directly, and seamlessly.

It starts speaking before it finishes thinking: As soon as it determines the first few words, it starts talking immediately. The rest of the sentence is processed while you are already listening.
It understands meaning, not just letters: Its tone shifts naturally based on the meaning and context of what it says, rather than flatly reading letters off a screen.
No error accumulation: Without multiple disconnected subsystems, small errors can no longer compound on top of each other to ruin the overall response.

Making AI Affordable for Khmer

Here is a hidden challenge in AI: Khmer is computationally expensive.

Because of how standard AI breaks down language (called "tokens"), processing a sentence in Khmer previously required significantly more computational power compared to English. This resulted in slower responses, higher server costs, and reduced memory capacity for the AI.

Before teaching Kiri-S2S to speak, we taught it to read Khmer efficiently.

We added 4,000 Native Khmer Tokens to its dictionary, which reduced the processing cost for Khmer by up to 80%. Now, processing Khmer costs almost the same as processing Englishmaking Real-Time Khmer Voice AI truly feasible for the first time!

To make this concrete, here is one everyday Khmer sentence "ខ្ញុំចង់រៀនភាសាខ្មែរឱ្យបានល្អជាងមុន ព្រោះវាជាភាសាកំណើតរបស់ខ្ញុំ។" ("I want to learn Khmer better, because it is my native language.") encoded by the tokenizers of today's leading AI models:

Tokenizer	Tokens for the Khmer sentence	Cost vs. English*
GPT-4 / GPT-3.5 (`cl100k`)	99	7.1×
Llama 3.1	99	7.1×
Mistral	83	5.9×
DeepSeek V3	80	5.7×
Qwen 3	78	5.6×
Gemma 2	63	4.5×
GPT-4o (`o200k`)	42	3.0×
Gemma 3	31	2.2×
Kiri-S2S	15	1.1×

* The same sentence in English is 14–15 tokens under every tokenizer shown. With Kiri-S2S's vocabulary, Khmer finally costs the same as English and every token saved is latency, context window, and compute returned.

How We Built It

We trained Kiri-S2S in 4 distinct phases:

Vocabulary: Teaching the AI to read Khmer efficiently.
Knowledge: Feeding it ~1 million clean Khmer web documents and hundreds of thousands of Khmer-English conversational pairs.
Listening: Teaching it to understand how humans converse in real-world scenarios.
Speaking: Teaching it to translate thoughts into natural-sounding voice output.

The final model contains 2.04 billion parameters, but it has been optimized so efficiently that the entire systemboth training and live inferencecan run on a single GPU card!

Current Limitations & What to Expect

Because Kiri-S2S is currently released as a Research Preview, there are still limitations that we are actively working to improve:

Occasional Mistakes & Hallucinations: Like any frontier AI model, Kiri-S2S may occasionally misinterpret queries or generate inaccurate information.
Background Noise Sensitivity: It performs best in quiet environments. Heavy background noise or overlapping voices can sometimes cause it to mishear input.
Complex Technical Jargon: Highly specialized technical terms or complex borrowed vocabulary may occasionally cause minor hesitations in response generation.

Help Us Improve: Your testing and feedback are critical to this process! If you encounter any bugs or unnatural responses, please let us know so we can refine Kiri and make it even smarter for everyone in Cambodia.

Looking Ahead

Kiri-S2S isn't just a brief tech demo it is the second milestone in a much larger mission.

We are Blizzer, a Cambodian AI startup. We hold a simple yet firm belief: For Cambodians, voice is the ultimate bridge to communicate with AI. Not typing. Not English. Native Khmer voice.

Our roadmap to that future is built on three clear pillars:

Step 1 KiriTTS: Giving Khmer a natural "Voice" to read text out loud.
Step 2 Kiri-S2S: Unlocking "Conversation" by building an AI that listens, understands, and responds in real time.
Step 3 AI Agents: Enabling "Actions" so AI can perform real-world tasks on your behalf.

Imagine calling a medical clinic to book an appointment, ordering food, or asking about insurance details fluently in Khmer without pressing buttons or switching to English for a single word.

The foundational technology to make that future a reality is already embedded right inside Kiri-S2S.

KiriTTS gave Khmer a voice.
Kiri-S2S is giving it conversation.
What comes next will give it the ability to take action.

And we are hard at work building all three for Cambodia!