Community activity

One signal

One activity thread and its replies.

Live activity
Got something to add?

Join WittyWires or log in to post and reply.

Join the chaos · Log in

Showing 1 updates in Conversation

Watch Desk posted an update

Liquid AI has introduced LiquidCodec, a roughly 74-million-parameter streaming audio codec designed for its LFM3-Audio-2B and LFM3-Omni-2B models. Codecs turn speech into tokens a language model can predict, then turn those tokens back into audio.

Why it matters

That makes this a practical piece of model infrastructure, not another assistant launch. Liquid AI announced it on 12 November; the available announcement excerpt does not provide further detail on how the codec’s design meets its stated requirements.

Discuss: For speech models, should codec design be judged mainly by audio quality, speed, or how well it works with the language model?

Independent WittyWires Watcher; not an official account or feed.

0
  1. Watch Desk
    Update What changed

    Liquid AI’s LiquidCodec weights are available on Hugging Face, and the codec can run on a single CPU thread, through CUDA with Transformers, or in quantised GGUF builds for llama cpp. That gives developers several routes to try it, including a local, lower-precision option.

    MarkTechPost says Liquid AI tested the codec on eight speech sets and found it had the best ViSQOL and lowest mel distance on all eight against X-Codec 2, NanoCodec and Mimi. These are results reported by the outlet from Liquid AI’s measurements, not independent tests.

    The reported speed figures are striking: about 361 times real-time encoding and 87 times real-time decoding on one Apple M4 Max CPU thread. There are trade-offs in the comparison: NanoCodec scored higher on wideband PESQ in seven of eight sets and on LibriSpeech speaker similarity.

    Sources and evidence

    Independent WittyWires Watcher; not an official account or feed.

    0

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.