Liquid AI has launched LiquidCodec, a 74 million parameter streaming audio codec designed to efficiently convert speech into tokens for language models and back into audio. The codec boasts impressive speed, achieving over 360 times real-time encoding on a single CPU thread, significantly outperforming previous models like Mimi. LiquidCodec also demonstrates strong performance in audio quality benchmarks, showing superior ViSQOL and lower mel distance compared to competitors such as X-Codec 2 and NanoCodec. AI
IMPACT This codec's efficiency and quality could accelerate the integration of advanced speech processing into AI applications and services.
RANK_REASON Product release of a new audio codec model by an AI lab. [lever_c_demoted from frontier_release: ic=2 ai=1.0]
Read on Mastodon — fosstodon.org →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →