ORI-Realtime 1.5 - Waterr.ai

ORI-Realtime 1.5 is Waterr’s end-to-end realtime interaction model: it listens, decides when to speak, thinks, and answers in a natural voice — over a single WebSocket connection. It is the same engine that runs live Waterr meetings, exposed directly to developers. Status: Research Preview. Access is waitlist-gated — request access.

Architecture

ORI-Realtime is a system, not a single network. Four layers run as one pipeline:

Layer	What it does
Semantic turn engine	Decides when the user has finished speaking using semantic voice-activity detection — not a fixed silence timeout. This is what makes barge-in and fast handoffs feel natural.
Streaming speech-to-text	Word-level streaming transcription, built on Kyutai’s open streaming STT foundation.
Reasoning layer	A swappable LLM that generates the response as the user is still being transcribed.
Streaming text-to-speech	Voice synthesis starts before the full response text exists, built on Kyutai’s open streaming TTS foundation (voice catalog).

The API surface follows OpenAI Realtime API event shapes — if you’ve built against a realtime speech API before, the event flow will look familiar. See the WebSocket reference.

Capabilities

Speech-to-speech conversation with streaming text transcripts of both sides
Natural interruption — barge-in is handled server-side by the semantic turn engine; just keep streaming microphone audio
9 preset voices (plus custom reference support for enterprise) — see voices
Steerable persona via plain-text instructions in session.update
Word-level transcript deltas for both user and model speech, with natural spacing

Numbers (preview, honest)

Metric	Value
First-audio latency	~1.2–1.5 s after end of user turn (improving; target < 1 s)
Audio format	Opus, 24 kHz mono, base64 over JSON (details)
Language	English (primary); other languages experimental
Concurrent sessions per key	2
Max session length	30 minutes
Pricing	Free during research preview; sessions are metered by duration

Limitations

Be aware of these before you build:

No client-side response.cancel. You cannot force-stop a response from the client; interruption happens server-side when the user speaks. This is a deliberate design of the turn engine, but it differs from OpenAI’s Realtime API.
Sessions are ephemeral. If the connection drops, a reconnect starts a fresh session — conversation history is not restored. Keep your own transcript if you need continuity.
Audio is Opus, not PCM16. Drop-in OpenAI Realtime clients need their audio encode/decode swapped — see audio.

Built in the open

The streaming STT and TTS foundations are open models released by Kyutai (CC BY 4.0); preset voices derive from the CSTR VCTK Corpus (CC BY 4.0). The turn engine, reasoning orchestration, serving stack, and this API are Waterr’s. Full attributions: voice attributions.

Quickstart

Key to first conversation in five minutes.

WebSocket reference

Every event, field, and close code.

​Architecture

​Capabilities

​Numbers (preview, honest)

​Limitations

​Built in the open