Noise Robust Keyword Spotting for Low-power Speech Communication Systems - with Application to Hearing Aids
Author
El-Sayed, Mohammad
Term
4. term
Education
Publication year
2016
Submitted on
2016-06-02
Abstract
Dette speciale undersøger brugen af keyword spotting (KWS) i lavenergi tale-kommunikationsenheder, med særligt fokus på høreapparater. Formålet er at muliggøre håndfri betjening via et stemmeaktiveret vågeord. Et letvægts KWS-rammeværk fra litteraturen udvælges med hensyn til de stramme beregnings- og energikrav i mobile enheder og implementeres. Systemets evne til at opdage mål-ord og afvise andre ord evalueres på virkelige høreapparatoptagelser foretaget af forfatteren. Forsøgene viser, at det valgte system er konkurrencedygtigt i forhold til samtidige KWS-metoder og overgår dem på nogle punkter. For at forbedre robustheden i støj introduceres to bag-end strategier til støjtilpasning; den ene strategi viser bedre ydeevne end både basissystemet og klassiske tale-forbedringsalgoritmer. Arbejdet demonstrerer potentialet for pålidelig, støjrobust stemmeaktivering i høreapparater med lavt strømforbrug.
This thesis investigates the use of keyword spotting (KWS) in low-power speech communication devices, with a focus on hearing aids. The goal is hands-free operation via a voice-activated wake word. A lightweight KWS framework from the literature is selected to meet tight computational and energy constraints and is implemented. The system’s ability to detect target keywords and reject non-keywords is evaluated using real-life recordings made with hearing aids. Experiments show that the chosen system is competitive with contemporary KWS approaches and outperforms them in some aspects. To enhance noise robustness, two back-end noise adaptation strategies are proposed; one of them achieves superior performance compared with the baseline KWS and with classical speech enhancement algorithms. The work demonstrates the feasibility of reliable, noise-robust voice activation in low-power hearing aids.
[This summary has been generated with the help of AI directly from the project (PDF)]
Documents