LiquidCodec
LiquidCodec is a streaming neural audio codec open-sourced by Liquid AI, with approximately 74M parameters. It adopts a low-frame-rate, multi-codebook design to provide efficient discrete tokenization for large audio models.
Timeline (last 90 days)
Liquid AI has released a streaming audio codec called LiquidCodec, with approximately 74M parameters.
LiquidCodec is the underlying component that provides audio encoding and decoding capabilities for the two audio models LFM3-Audio-2B and LFM3-Omni-2B, and is not a standalone product aimed at end users.
LiquidCodec runs at a frame rate of 12.5 Hz and uses 12 codebooks to represent audio information.
According to officially disclosed data, LiquidCodec can achieve an encoding speed of about 361 times real-time on a single CPU thread.
LiquidCodec's encoding requires less than one second of computation to process one minute of audio on a single CPU thread, and does not rely on GPU acceleration.
A 12.5 Hz frame rate means one time step every 80 milliseconds, a typical low-frame-rate streaming configuration.
Having multiple models share the same codec gives audio input and output a unified token space, making it easier to do speech understanding, generation, and full-duplex dialogue within the same framework, and saving the cost of training a separate tokenizer for each model.
361 times real-time measures throughput (the computation time needed to process a unit duration of audio) and is not equivalent to interaction latency; true streaming latency also depends on whether the algorithm needs to see future context, the chunk window size, and the extra delay introduced by the codec round trip.
Currently available public information has not yet disclosed LiquidCodec's subjective audio quality scores, horizontal comparisons with other codecs, or its latency performance under real network jitter conditions.
All Facts (9)
Currently available public information has not yet disclosed LiquidCodec's subjective audio quality scores, horizontal comparisons with other codecs, or its latency performance under real network jitter conditions.
50%UnverifiedLiquidCodec is the underlying component that provides audio encoding and decoding capabilities for the two audio models LFM3-Audio-2B and LFM3-Omni-2B, and is not a standalone product aimed at end users.
50%UnverifiedLiquidCodec runs at a frame rate of 12.5 Hz and uses 12 codebooks to represent audio information.
50%UnverifiedAccording to officially disclosed data, LiquidCodec can achieve an encoding speed of about 361 times real-time on a single CPU thread.
50%UnverifiedLiquidCodec's encoding requires less than one second of computation to process one minute of audio on a single CPU thread, and does not rely on GPU acceleration.
50%UnverifiedA 12.5 Hz frame rate means one time step every 80 milliseconds, a typical low-frame-rate streaming configuration.
50%UnverifiedHaving multiple models share the same codec gives audio input and output a unified token space, making it easier to do speech understanding, generation, and full-duplex dialogue within the same framework, and saving the cost of training a separate tokenizer for each model.
50%Unverified361 times real-time measures throughput (the computation time needed to process a unit duration of audio) and is not equivalent to interaction latency; true streaming latency also depends on whether the algorithm needs to see future context, the chunk window size, and the extra delay introduced by the codec round trip.
50%UnverifiedLiquid AI has released a streaming audio codec called LiquidCodec, with approximately 74M parameters.
50%