Speech to Decisions with Granite

Tap to start talking

Your words are judged as you say them

00:00

Microphone

Transcript Granite Switch: ASR Encoder

Your words will appear here as you speak…

The speaker is talking about Granite Switch: Base Model

—

Safety check Granite Switch: Guardian Adapter

Yes/No questions (edit the list to customize) Granite Switch: Decisions Adapter

Talk into your microphone. One model, Granite Switch 4.1 8B, does everything from one endpoint: its ASR encoder turns your speech into text, and about twice a second it judges what you're saying three ways: the base model names what you're talking about, the guardian adapter screens for harmful content, and the decisions adapter answers every yes/no question.

Every one of those judgments starts from the same words, so the endpoint reads them once and every request reuses that KV cache. That's why a whole board of answers comes back in a fraction of a second.

After 0.5 s of silence the board goes blank until you speak again. Your audio is sent to this Space and on to the Granite endpoint.

On the board:

What the speaker is talking about, in 1 to 5 words
A safety check for harm, profanity and violence
Yes/no questions you can edit before you talk: 10 to start, up to 50 in all