Liquid AI’s LiquidCodec converts 24 kHz mono speech into 12 tokens per frame at 12.5 frames per second and reconstructs audio with roughly 74 million parameters. Independent codebooks enable parallel prediction, and the codec powers LFM3 audio and multimodal tokenization. The company reports encoding at about 361 times real time and decoding at 87 times real time in FP32 on one M4 Max CPU thread. These speech-only results do not establish music performance, and published weights use the LFM Open License.
Back to Tech
LiquidCodec offers a roughly 74M-parameter streaming speech codec
It targets low-latency speech tokenization; music and environmental-audio quality have not been evaluated.