Liquid AI has published LiquidCodec, the in-house audio codec inside the LFM3-Audio-2B and LFM3-Omni-2B models it released at the end of October. A codec like this is the part of a voice model that turns speech into tokens a language model can predict, and turns predicted tokens back into sound. The weights are on Hugging Face.
The spec sheet
| LiquidCodec | |
|---|---|
| Parameters | about 74M |
| Audio | 24 kHz mono speech |
| Frame rate | 12.5 Hz, one frame every 80 ms |
| Tokens per frame | 12, one per codebook |
| Bitrate | about 1.65 kbps |
| Streaming | encoding and decoding |
| Training data | 264k hours across 107 languages |
The model card says it was trained and evaluated on speech only; quality on music, singing or other sounds has not been tested. It loads natively in Transformers through LiquidCodecModel, with a LiquidCodecStreamingCache for chunked synthesis.
Why Liquid built its own
Liquid’s first audio model, LFM2-Audio-1.5B, used the Mimi codec. LFM2.5-Audio replaced only Mimi’s decoder, which Liquid says made decoding 8 times faster. LiquidCodec replaces the tokenizer too. Instead of residual vector quantization, where each codebook depends on the one before, it uses grouped finite scalar quantization. The 12 codebooks in a frame are independent, so a speech model can predict them all in parallel. The encoder mirrors the decoder and is fully causal, which suits streaming.
On speed, Liquid reports that on a single CPU thread at FP32 on a MacBook Pro with an M4 Max, LiquidCodec encodes at about 361 times real time and decodes at about 87 times. Its table puts Mimi at a real-time factor of 0.036 to encode and 0.033 to decode, against 0.0028 and 0.0115 for LiquidCodec. These are Liquid’s measurements.
The license
This is open weights, not open source. The LFM Open License v1.0 allows commercial use only while your company’s annual revenue stays under $10 million; above that you need a paid license. It is the same deal we flagged for the LFM3 family, and similar to Fish Audio’s revenue-capped Drama 3.
Why this matters
The codec sets a floor on how fast a voice model can talk back, because every spoken token passes through it. A 74M codec that streams and decodes faster than real time on one CPU thread, by Liquid’s numbers, is a useful part for local voice agents. Check the revenue cap before you build a product on it.
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment