A speech large language model trained on beamformed multi-channel audio performs directional speech recognition and source localization across 12 discrete angles on simulated smart glasses data.
We propose two key techniques to infuse directional knowledge: serialized directional output training (S-DOT) and contrastive direction data augmentation (CDDA)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
eess.AS 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Thinking in Directivity: Speech Large Language Model for Multi-Talker Directional Speech Recognition
A speech large language model trained on beamformed multi-channel audio performs directional speech recognition and source localization across 12 discrete angles on simulated smart glasses data.