LIVE

Liquid AI Debuts LiquidCodec to Accelerate Streaming Speech Processing

🧠 AI Analysis 2 min read 👁 11

Liquid AI has launched LiquidCodec, a 74-parameter streaming audio codec designed to convert speech into language model tokens efficiently. The new tool aims to improve speed and audio quality over prior systems like Mimi.

🧠AI Analysis

Key points

  • LiquidCodec features a 74-parameter streaming speech tokenizer architecture.
  • It replaces Kyutai's Mimi codec, utilizing grouped finite scalar quantization for parallel codebook prediction.
  • Self-reported tests show high efficiency, running on a single CPU thread at hundreds of times real-time speed.

Liquid AI has unveiled LiquidCodec, a streaming speech tokenizer operating at approximately 74 million parameters that underpins its LFM3-Audio-2B and LFM3-Omni-2B models. The system converts spoken audio into tokens that language models can easily predict, then decodes those tokens back into sound. According to the company, the development focused on generating predictable tokens and enabling rapid conversion back to speech, replacing Kyutai's Mimi codec used in previous iterations.

The architecture relies on an encode, quantize, and decode pipeline utilizing symmetric blocks from the LFM3 framework with a fully causal encoder suited for streaming. Instead of residual vector quantization, Liquid AI employs grouped finite scalar quantization, allowing all twelve codebooks to be predicted in parallel. The training process spanned two stages across hundreds of thousands of hours of speech in multiple languages, followed by fine-tuning with adversarial critics to eliminate metallic artifacts.

Self-reported benchmarks tested on a single CPU thread of an Apple M4 Max laptop indicate high efficiency, achieving roughly 361 times real-time encoding and 87 times real-time decoding. The company also states that LiquidCodec achieved superior ViSQOL scores and lower mel distance across eight test sets compared to alternatives like X-Codec 2, NanoCodec, and Mimi. Model weights are currently accessible on Hugging Face.

Source MarkTechPost — marktechpost.com/2026/11/12/liquid-ai-liquidcodec-streaming-audio-c… ↗

Contact the newsroom

Got a tip, a proposal or a partnership question? Write to us and we will reply shortly.