gatilin/lfm3-omni-audio-encoder

LFM3 Omni audio encoder and projector

Community extraction of model.audio_encoder.* and model.audio_projector.* from LiquidAI/LFM3-Omni-2B, revision 21d75123f86c1f7ffb4bef1ea202710c4e7ebe67. It provides the speech input encoder and language-space projector; the language model, output decoder and LiquidCodec are not included. Tensor keys, dtypes, shapes and values are preserved.

The included audio feature extractor expects mono 16 kHz audio. This is distinct from the separate 24 kHz LiquidCodec. See extraction.json, tensor_mapping.json and validation.json for provenance and test limits.

Loading

Install PyTorch, Transformers 5.21.1, accelerate, safetensors and NumPy. Use component.load_model(root) and component.load_feature_extractor(root) after downloading the repository. The feature extractor returns log-mel input_features and feature_attention_mask. Feed the feature tensors and their valid lengths to model.model.audio_encoder, then pass valid encoded features to model.model.audio_projector. A single audio input returns one output token for each eight mel frames.

The CPU test is a synthetic signal smoke test; it is not a speech quality benchmark, full-model equivalence test or GPU validation.

License

Distributed under the original LFM Open License 1.0, including its commercial-use conditions. Source code copyright and licensing notices are retained in the bundled files.