Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)

Liquid AI Releases LiquidCodec: Decodes Speech 2.9x Faster Than Mimi

A voice assistant on your phone won't respond quickly if audio takes too long to turn into tokens and back. Liquid AI has released LiquidCodec, a 74M-parameter codec that decodes one second of audio in 0.0115 seconds. That is 2.9 times faster than Mimi (0.0329 s). Quality is also higher on the ViSQOL metric on LibriSpeech: 4.319 versus 4.031 for Mimi. If latency is the bottleneck in your voice stack, the codec is worth testing on your own hardware.

AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
Liquid AI Releases LiquidCodec: Decodes Speech 2.9x Faster Than Mimi
Source: TG: @ai_machinelearning_big_data (Machinelearning). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @ai_machinelearning_big_data (Machinelearning) ↗
Editor-in-Chief
Key metric #1Encoding on a single M4 Max thread runs 361 times faster than real time
Key metric #2The 4-bit decoder is 3.5 times smaller, with ViSQOL almost unchanged at 4.292
Key metric #3LFM Open v1.0 license, weights available on Hugging Face

Imagine a voice assistant that runs directly on your phone or laptop, with no cloud. You say: "Set a timer for ten minutes." The language model doesn't hear audio.

It reads discrete tokens, which are sequences of numbers. A codec has to turn speech into those numbers and assemble the response back into sound. If the codec is slow, the assistant answers with a pause, even when the model itself is fast.

Liquid AI has released LiquidCodec, a 74M-parameter codec, as open source. It decodes one second of audio in 0.0115 seconds, while Mimi, the codec of the previous LFM2-Audio, takes 0.

0329 seconds. That is 2.9 times faster (editorial calculation).

If you build voice features for devices, this is a direct reason to test the codec on your own data.

What the codec does and how this one differs

The encoder slices audio into frames and records each frame as a set of numbers drawn from predefined lookup tables called codebooks. The model reads these numbers and predicts the continuation. The decoder reconstructs audible speech from the predicted numbers. LiquidCodec underlies LFM3-Audio-2B and LFM3-Omni-2B, which came out in late October.

LiquidCodec produces 12.5 frames per second at a 24 kHz audio sampling rate. Its encoder is fully causal: it looks only at the past and does not wait for the end of a phrase. That makes the codec suitable for streaming from a microphone, where the recording is not accumulated in advance.

Mimi uses residual vector quantization, in which codes refine each other in a chain. LiquidCodec uses grouped finite scalar quantization and twelve codebooks. The language model can predict them in parallel rather than guessing them one after another. By my calculation, at 12.5 frames per second that comes to about 150 tokens per second of audio. The decoder outputs spectral coefficients directly and avoids costly learned upsampling.

What the measurements showed

All figures come from Liquid AI's own post and blog. There are no independent checks yet.

  • On a single CPU thread of a MacBook Pro with an M4 Max, encoding runs 361 times faster than real time and decoding runs 87 times faster.
  • The decoder spends 0.0115 seconds per second of audio. For Mimi it is 0.0329, for X-Codec 2 it is 0.0379, and for NanoCodec it is 0.2543.
  • ViSQOL, an automatic measure of how close reconstructed speech is to the original, came to 4.319 on LibriSpeech. NanoCodec scored 4.219, X-Codec 2 scored 4.119, and Mimi scored 4.031.
  • On ViSQOL and mel-spectrogram distance, the codec is the best on all eight datasets. The languages are English, Chinese, German, French, Japanese and Korean.
  • Speech synthesis models trained on LiquidCodec tokens achieved the best naturalness on all eight datasets and the best intelligibility on four.
  • The decoder with 4-bit quantization is 3.5 times smaller. Quality barely drops: ViSQOL is 4.292.

What to do with this

Two things are useful for developers. First, a lightweight codec with open weights can be taken on its own, for example for your own speech model, and not only together with LFM3. Second, the 4-bit decoder leaves headroom in memory for mobile and embedded scenarios.

There are caveats too. Speed was measured on a powerful Apple processor, and numbers on a budget phone will differ. ViSQOL and mel-spectrograms are automatic measures and do not replace listening tests with real people. Synthesis intelligibility is best on only four of the eight datasets. The LFM Open v1.0 license is not one of the most familiar open licenses, so for a commercial product it is better to read its terms before integration.

If your voice stack is currently bottlenecked by decoding latency, LiquidCodec is worth testing on your own data and your own hardware. The weights and the blog post are already published.

Story

Key Questions Answered

What the codec does and how this one differs?

The encoder slices audio into frames and records each frame as a set of numbers drawn from predefined lookup tables called codebooks. The model reads these numbers and predicts the continuation. The decoder reconstructs audible speech from the predicted numbers. LiquidCodec underlies LFM3-Audio-2B and LFM3-Omni-2B, which came out in late October. LiquidCodec produces 12.5 frames per second at a 24 kHz audio sampling rate.

What the measurements showed?

All figures come from Liquid AI's own post and blog. There are no independent checks yet. On a single CPU thread of a MacBook Pro with an M4 Max, encoding runs 361 times faster than real time and decoding runs 87 times faster. The decoder spends 0.0115 seconds per second of audio. For Mimi it is 0.0329, for X-Codec 2 it is 0.0379, and for NanoCodec it is 0.2543. ViSQOL, an automatic measure of how close reconstructed speech is to the original, came to 4.319 on LibriSpeech.

What to do with this?

Two things are useful for developers. First, a lightweight codec with open weights can be taken on its own, for example for your own speech model, and not only together with LFM3. Second, the 4-bit decoder leaves headroom in memory for mobile and embedded scenarios. There are caveats too. Speed was measured on a powerful Apple processor, and numbers on a budget phone will differ. ViSQOL and mel-spectrograms are automatic measures and do not replace listening tests with real people.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…