SAV-BandNet

A six-channel, visually conditioned target-speech extractor using 1 s windows, 0.4 s emissions and 0.2 s lookahead. The image frontend is TalkNet; spatial features come from visually conditioned IPDNet2. Candidate streams come from an adapted DAVE separator. Target selection combines audiovisual scores, recurrent history and speaker identity memory.

This release contains the earlier TalkNet extraction checkpoint trained on 4,096 scenes. It is not the later 32,768-scene paper model, and historical CMLR or full-training-set metrics should not be attributed to it. The original head hash is 55505a92fcb523457790ff0d9f1aedb1aab4986eaf57f0d85c6e3b357fa09b59. Component hashes and lineage are in PROVENANCE.json. Optimizer states and historical parent checkpoints are omitted.

Use

Download the bundle and implementation.tar.gz. Extract the implementation, install its dependencies following docs/setup.md, and set SAV_BUNDLE to the downloaded bundle directory (or pass --bundle). The dataset is SAVE.

python -m sav_bandnet infer --data /path/to/SAVE --bundle /path/to/bundle --output /path/to/results

Inputs are synchronized 16 kHz six-channel audio and 25 fps 96×96 target-mouth crops with explicit validity. The reference geometry is a six-microphone circular array of radius 3.5 cm. Different arrays, microphone ordering or audiovisual synchronization require validation. The code emits CPU-available waveform chunks; this is not a measured physical capture-to-playback latency guarantee.

RELEASE_VERIFICATION.json records parity and bounded training smoke checks. These checks do not establish convergence of the consolidated training command or real-world performance. Candidate alignment uses overlapping waveform correlation; speaker embeddings correct selection scores.

Sources and terms

The implementation derives from DAVE/TIGER, TalkNet and IPDNet2. Retained source notices accompany the implementation. Component code licenses do not automatically license upstream pretrained weights or source media. No blanket commercial-use or redistribution-rights assertion is made for all components. Consult the upstream sources and THIRD_PARTY_NOTICES.md.