From the source
Liquid AI has built LiquidCodec, an in-house streaming audio codec that turns speech into tokens for its LFM3 audio models and converts them back into audio.
- The company says the roughly 74M-parameter codec matches similarly sized models on reconstruction quality and compression while beating them on speech generation and token-to-waveform decoding speed.
- Liquid AI previously relied on the third-party codec Mimi in LFM2-Audio-1.5B, then replaced only its detokenizer in LFM2.5-Audio-1.5B for 8x faster detokenization, but the third-party tokenizer still limited audio-token quality and architecture choices.
- LiquidCodec swaps Mimi's residual vector quantization for grouped finite scalar quantization, which reduces sequential dependencies between codebooks and allows parallel prediction of audio tokens.
- The codec keeps a 12.5 Hz frame rate and models 24 kHz waveforms, with a symmetric encoder and decoder and a fully causal encoder for real-time, low-latency streaming.
- Training ran in two multilingual stages, the first on 264,000 hours across 107 languages and the second fine-tuned only on high-fidelity waveforms.





