Oriel — AI & technology worth noticing, every day.

LiquidAI/Liquid-Codec

AI model · https://huggingface.co/LiquidAI/Liquid-Codec

First seen by Oriel on .

AI summary

LiquidCodec is a compact, streaming speech codec whose parallelizable GFSQ tokens deliver high reconstruction and TTS quality with fast CPU inference and quantized edge deployment.

  • LiquidCodec uses a fully causal LFM3 encoder and decoder with GFSQ, producing 12 tokens per 12.5 Hz frame at approximately 1.65 kbps.
  • The approximately 74M-parameter codec processes 24 kHz mono speech, achieving about 361× real-time encoding and 87× decoding on a single Apple M4 Max CPU thread.
  • Across eight reconstruction test sets, LiquidCodec achieved the best ViSQOL and lowest mel distance, including with a Q4_0 decoder.
  • Matched zero-shot TTS models using LiquidCodec tokens achieved the highest UTMOSv2 on all eight test sets and best WER/CER on four.
  • Decoder-only quantization-aware training enables Q4_0 deployment with a 3.5× smaller footprint and minimal quality loss; music and environmental sound were not evaluated.

#audio codecs #speech synthesis #quantization #streaming inference

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 154, latest observation ; +46 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 9,378, latest observation ; 0 over observed period since

Aliases

SourceExternal IDURL
Hugging Face ModelsLiquidAI/Liquid-Codechttps://huggingface.co/LiquidAI/Liquid-Codec Read discussion