Liquid AI Releases LiquidCodec, a 74M-Parameter Streaming Audio Codec
Summary
- Liquid AI has released LiquidCodec, a 74M-parameter streaming speech tokenizer and audio codec designed for models like LFM3-Audio-2B.
- It uses grouped finite scalar quantization (GFSQ) to enable independent and parallel token prediction across 12 codebooks.
- LiquidCodec achieved competitive benchmark scores, including top ViSQOL and mel distance metrics across eight test sets compared to alternatives like Mimi.
- The codec supports deployment via Hugging Face Transformers and GGUF builds for llama.cpp, running efficiently on CPU threads and CUDA GPUs.
Why it matters
Efficient streaming audio codecs are critical for real-time voice AI and multimodal language models. LiquidCodec's high processing speed and parallel token prediction make edge deployment and speech generation more practical for developers.