Playground

A real WebRTC call through the same /v1/realtime/calls endpoint your apps use: your voice, then ASR, the LLM and TTS, with noise reduction and turn detection on the server.

No call yet

Talk to your cascade

Not connected

No conversation yetStart a call and say hello.

Call settings

Sent as the session for the next call.

Instructions

Voice

Noise reduction

Providers for this call

ASR

LLM

TTS

Provider's own voice

Remembered in this browser.

Turn detection

Events

Calls & cost

What your calls cost. Every call is metered as it runs: what the cascade actually spent, and what the same conversation would have cost on a speech-to-speech Realtime model.

Calls

Call	Length	Turns	Reply time	Cascade	Realtime	Saved	Cascade vs realtime

Settings

Services, session defaults and prices. Saved to ; new sessions pick changes up within a second.

Recognition

Keeps short or noisy turns from coming back in the wrong language.

Caller languagesISO codes, main one first, e.g. ar, en. A turn heard in any other language is transcribed again in the conversation's language.

Give the recognizer the conversation so farKeeps the language, names and spellings consistent from turn to turn.

Replies

How replies are requested from /chat/completions.

Reasoning effortFor reasoning models (gpt-5, o-series, gpt-oss). Thinking happens before the first word, so less is faster.

Reply length fieldHow the reply token cap is sent. Newer OpenAI models reject max_tokens.

Extra request fields (JSON)Merged into every chat completion, e.g. {"temperature": 0.6}.

Tell the model its replies are spokenAdds a short note so answers come out as plain sentences, without markdown.

Providers

Named services the stages below and test calls can pick. The openai kind covers any OpenAI-compatible API: OpenAI, Groq, DeepSeek, Together, OpenRouter, vLLM, Ollama. Keys can read the environment, e.g. ${GROQ_API_KEY}.

Voices

Realtime voice names your clients send, mapped to your speech service's voices. One per line: marin=coral.

Voice map

Default voice

PCM sample rateWhat your TTS returns for pcm.

Buffer before speaking (ms)Audio held back so speech doesn't stutter.

Speaking style

Connect

The server speaks the OpenAI Realtime API (GA and beta), so existing apps only need a new base URL.

Ahleen

In the model provider, choose OpenAI-compatible and set:

Base URLModelgpt-realtimeAPI keyAny client key

Endpoints

Everything the Realtime API offers, at this address.

Playground

Talk to your cascade

Call settings

Events

Calls & cost

Calls

Settings

Recognition

Replies

Providers

Voices

Session defaults

Turn detection

What counts as speech

Cascade prices

Compared against

Access

WebRTC & history

config.yaml

Connect

Ahleen

Endpoints

OpenAI SDK (Python)

Browser (WebRTC)