🤖 🌟 Liquid AI releases speech codec for its LFM3 audio models

LiquidCodec is a 74M-parameter audio codec (LiquidAI/Liquid-Codec repository on Hugging Face, LFM Open v1.0 license). The codec outputs 12.5 frames per second with 12 tokens each (~1.65 kbps); instead of RVQ, as used in Mimi, GFSQ is applied — 12 codebooks per frame are predicted in parallel.

🌍 GFSQ reduces sequential autoregression steps for audio-LLMs. On a single CPU thread of an Apple M4 Max, decoding is 87 times faster than real-time, and encoding is 361 times faster; according to ViSQOL 4.319 on LibriSpeech, the codec leads in all eight tests.

👤 BF16 weights via Transformers (trust_remote_code=True) and Q4_0, Q4_K_M, Q8_0 decoders in GGUF for llama.cpp allow building a voice assistant on a laptop and phone. The figures are currently vendor-reported (single CPU setup), and the license text is not fully disclosed.

Source 1: https://www.liquid.ai/blog/liquidcodec Source 2: https://huggingface.co/LiquidAI/Liquid-Codec