Skip to content

Kimi-Audio

An open audio foundation model for speech and audio tasks

Visit official site

What Kimi-Audio is

The model is described as an open-source audio foundation model for speech recognition, audio conversation and related audio tasks, pretrained on more than 13 million hours of diverse audio and text data. It combines hybrid audio input with an LLM core and a flow-matching block-wise streaming decoder.

What you can do with it

  • Recognise speech in audio input
  • Answer questions about audio content
  • Generate captions for audio content
  • Handle text and audio input at the same time

Model notes

  • Pretraining uses more than 13 million hours of audio and text data
  • A hybrid audio input and LLM core processes both modalities
  • A flow-matching block-wise streaming decoder supports low-latency generation
  • Code, model checkpoints and an evaluation toolkit are released

Getting started

  • Download the model and code from the GitHub page
  • Install the required dependencies and check the environment
  • Load the model and set the sampling parameters
  • Prepare audio input or dialogue information

Pros & cons

✓ What we like

  • Speech recognition, audio QA and audio captioning in one model
  • Block-wise streaming decoder targets low-latency generation

! What to watch out for

  • The described workflow starts from downloading code and checkpoints
  • Audio input or dialogue information is prepared before inference

FAQ

How large is the pretraining data?

The material states more than 13 million hours of diverse audio and text data.

What is released with the model?

The feature list names code, model checkpoints and a comprehensive evaluation toolkit.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review