Liquid AI Releases 74M-Parameter Streaming Audio Codec
Liquid AI has released LiquidCodec, a streaming speech tokenizer designed for the company’s LFM3-Audio-2B and LFM3-Omni-2B models. The codec has about 74 million parameters, including 73.6 million listed on its model card. It processes 24 kHz audio at 12.5 Hz, producing 12 discrete token streams, with each 80-millisecond frame represented by 12 tokens.
LiquidCodec uses a causal encoder and grouped finite scalar quantization, allowing its 12 codebooks to be predicted in parallel. Its decoder generates spectral coefficients directly rather than relying on learned transposed-convolution upsampling. Training used 264,000 hours of speech in 107 languages, followed by fine-tuning on 55,000 hours of full-band speech and a quantization-aware training stage for edge deployment.
In Liquid AI’s tests on an Apple M4 Max, encoding reached about 361 times real time and decoding about 87 times real time on one CPU thread. Across eight test sets, the codec recorded the best ViSQOL and mel-distance results among the compared systems, while NanoCodec led PESQ on seven sets. Weights are available on Hugging Face, with CUDA, Transformers and llama.cpp deployment options.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.