Abstract
Acoustic environments containing multiple simultaneous speakers present challenges for conventional noise cancellation and transparency technologies, which often fail to isolate specific audio based on semantic relevance. Traditional methods, such as directional beamforming or spectral profiling, cannot distinguish between sources based on the topical content of speech.
An adaptive semantic audio filtering method is disclosed to address these limitations. Ambient audio is captured and decomposed into discrete streams using neural blind source separation. Each stream is transcribed via automated speech recognition and converted into a context vector. These vectors are compared against a user-defined target semantic profile using cosine similarity scoring. Based on these scores, dynamic gain weights are applied to the separated streams to amplify relevant content while suppressing background noise. A circular look-ahead buffer is utilized to synchronize processing latency with the audio output. This technology enables the selective isolation of audio sources based on intent, improving clarity in complex acoustic environments.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Yakar, Tamar and Labzovsky, Ilia, "Adaptive Semantic Audio Filtering via Neural Source Separation and Intent-Based Vector Scoring", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9894