Live intelligence

Latest AI updates

2 updates
Models Editor selected

Liquid AI Releases LiquidCodec, a 74M-Parameter Streaming Audio Codec

Liquid AI has released LiquidCodec, a ~74M-parameter streaming speech tokenizer that powers its LFM3-Audio-2B and LFM3-Omni-2B models. It converts 24 kHz audio into 12 parallel token streams at 12.5 Hz using grouped finite scalar quantization and causal LFM3-based encoder/decoder blocks. In self-reported benchmarks, it claims the best ViSQOL, mel distance and TTS UTMOSv2 across eight test sets versus Mimi, NanoCodec and X-Codec 2, while running about 361x real-time encoding on a single CPU thread. Weights are available on Hugging Face with Transformers support and GGUF builds for llama.cpp.

Why it matters: This is a concrete new model release with strong efficiency and quality claims, plus practical deployment options for developers building streaming speech and TTS systems.

Liquid AIaudio codecsspeech processing
Research

DNN-Brain Alignment in Speech Processing Under Audiovisual Noise

This study examines how representations in a pretrained automatic speech recognition model align with human brain responses during acoustic-to-linguistic speech processing. Using neural encoding models and representational similarity analysis, the authors compare brain responses to acoustic and linguistic features with DNN embeddings before and after fine-tuning on audiovisual noisy data. Shallow layers show greater similarity to acoustic-related neural responses, while deeper layers align more closely with linguistic responses. Fine-tuning improves this alignment by enhancing noise processing in shallow layers and refining linguistic representations in deeper layers.

DNN-brain alignmentspeech processingaudiovisual noise