Liquid AI details its LiquidCodec audio tokenizer

Liquid AI has published a technical write-up of LiquidCodec, the streaming audio codec that underpins LFM3-Audio-2B and LFM3-Omni-2B. Both models were released at the end of October as part of the LFM3 family, which we covered earlier.

A codec turns speech into tokens that a language model can predict, then turns those tokens back into audio. Liquid's earlier audio models relied on Kyutai's Mimi codec (~96M parameters, 1.1 kbps). In LFM2.5-Audio-1.5B, Liquid replaced only Mimi's detokenizer, which made detokenization 8× faster. The company says keeping the third-party tokenizer capped both token quality and its control over the architecture.

LiquidCodec, which is built on LFM3 blocks, drops Mimi's residual vector quantization (RVQ) in favour of grouped finite scalar quantization (GFSQ). In RVQ, each codebook depends on the one before it. GFSQ largely removes that sequential dependency, so Liquid can predict several codebooks in parallel using MaskGIT-style methods. The encoder and decoder are exactly symmetric, and the encoder is fully causal, which suits low-latency streaming. Rather than upsampling with learned transposed convolutions, the decoder predicts spectral coefficients directly. The codec keeps a 12.5 Hz frame rate (one frame every 80 ms) and models 24 kHz audio.

Liquid says the ~74M-parameter codec matches similarly sized codecs on reconstruction quality and compression, and beats them on speech-generation and decoding speed. Benchmark figures were not in the material we reviewed. Training ran in two stages. The first used 264,000 hours of speech in 107 languages, and the second fine-tuned on high-fidelity audio only.