GPT-Audio Audio Understanding

Voice & speech playground

Hydrating page from local storage...

Ask questions about WAV or MP3 recordings, transcribe, translate, and summarize audio with text responses and follow-up turns.

Microphone access requires HTTPS or localhost and your permission. Audio is sent to the selected provider only when you submit a request or start a live session. Use headphones for live conversations.

GPT-Audio-1.5 accepts text and audio through /chat/completions. Each turn is a bounded request. Generated speech is AI-generated. Streaming displays incoming text and buffers audio for playback after completion.

Find models in Microsoft Foundry. Foundry audio generation guide

Connection

Use an OpenAI endpoint or a Foundry resource with the selected speech deployment.

Endpoint

Model / deployment name

Authentication Mode

API Key

Advanced

Speech controls

Maximum output tokens

Up to 16,384 output tokens. The model has a 128,000-token context window.

Temperature

Stream Include conversation history

System / delivery instructions

Describe tone, pace, language, or the task in natural language. Speech style and speed are prompt instructions. History replays previous input audio and assistant transcripts, up to 12 turns and 25 MB total input audio.

Enable local function-calling demo

lookup_demo_order returns fictional statuses for DEMO-100 and DEMO-200. It has no network access. Tool calls and results are inspectable in Payloads. A run can make up to three billable model requests.

Ready. Configure your connection to begin.

Audio input

Audio uploads are limited to 25 MB in this playground; microphone recordings stop at five minutes. GPT-Audio accepts WAV and MP3. Audio remains local until you submit. Choose Next turn to clear the composer and current recording while retaining enabled conversation context.

Text message (optional)

0 / 60,000 characters. 0 previous turns will be included.

Provider usage is billable. No estimated price is assumed; returned usage is available in Payloads.

Audio payload bytes are summarized to keep previews and archives bounded. Replace base64 placeholders when using generated code. Provider usage, completion reasons, and all tool rounds are retained in Payloads. Exact caption timings are unavailable.

Upload or record audio, add a prompt, and submit to see the model's response here.