gatilin/lfm3-audio-2b-encoder

LFM3-Audio-2B audio encoder and projector

Community extraction of encoder.* and mlp.* from LiquidAI/LFM3-Audio-2B, revision 2d6c1f99c129b3ef07a8d168bc0ff08e2bdb93f1. It contains the 17-layer audio input encoder and language-space projector. The language model, speech output decoder and LiquidCodec are not included. The 176 stored tensors retain source names, shapes, dtypes and values.

Four source tensor byte samples differ from the corresponding LFM3-Omni audio extraction; similar architecture does not mean identical weights. See contrast_with_omni.json for the sampling limits. component.py applies the original implementation's checkpoint key conversions during loading, without changing the stored tensor names.

Loading and validation

Install PyTorch, Transformers 5.21.1, accelerate, safetensors and NumPy. Use component.load_model(root) and component.load_feature_extractor(root) after downloading this repository. The feature extractor accepts mono 16 kHz audio and returns input_features and feature_attention_mask. Feed the features and valid lengths to model.model.audio_encoder; feed valid encoder outputs to model.model.audio_projector.

See validation.json for strict-loading and CPU smoke tests using a synthetic sine wave and a publicly released generated speech waveform. These are component checks, not a speech-recognition quality benchmark, full-model equivalence test or GPU validation. Source ranges, tensor metadata and hashes are in tensor_mapping.json and extraction.json.

LiquidCodec reuse

The official LFM3-Audio-2B and LFM3-Omni-2B audio_tokenizer/model.safetensors files have the same size and LFS SHA-256. The previously published LFM3-Omni LiquidCodec contains that independently verified file and can be reused. See liquidcodec_comparison.json for the pinned source revisions and digest.

License

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