Liquid AI unveils liquidcodec, a 74m-parameter streaming audio tokenizer

Liquid AI has introduced LiquidCodec, a new streaming speech tokenizer with approximately 74 million parameters, which powers its LFM3-Audio-2B and LFM3-Omni-2B models. This codec converts speech into predictable tokens for language model processing and back into audio. It replaces Mimi, which was used in the LFM2-Audio-1.5B model, and is available on Hugging Face. LiquidCodec compresses 24 kHz audio into 12 token streams, with each 80 ms frame yielding 12 discrete tokens. The pipeline involves encode, quantize, and decode stages, utilizing symmetric halves from the LFM3 architecture. It employs grouped finite scalar quantization (GFSQ) instead of residual vector quantization (RVQ), allowing parallel prediction via MaskGIT-style transformers. Performance benchmarks show significant efficiency gains; encoding achieved a real-time factor (RTF) of 0.0028, substantially lower than Mimi's 0.0361. The decoder RTF was 0.0115, compared to Mimi's 0.0329. Training involved two phases: one using 264,000 hours across 107 languages with waveform and spectral losses, and a second fine-tuning on 55,000 hours of full-band speech with adversarial critics to mitigate metallic artifacts. For deployment, Liquid AI offers Q4_0, Q4_K_M, and Q8_0 GGUF builds for llama.cpp, with the Q4_0 quantization shrinking the decoder footprint by 3.5x while minimally impacting LibriSpeech ViSQOL (from 4.319 to 4.292).