Transformer-Based Turkish Keyword Spotting in Noisy Environments Gürültülü Ortamlarda Dönüştürücü Mimari ile Türkçe Anahtar Kelime Tespiti
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636921
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: test-time adaptation, transformer architecture, Turkish keyword spotting, voice recognition
- Süleyman Demirel Üniversitesi Adresli: Evet
Özet
This paper presents a comprehensive evaluation of transformer-based approach for Turkish keyword spotting under challenging conditions. A baseline transformer was compared against three improved variants: KWTm1, which incorporates increased warm-up duration, added dropout, and decreased label smoothing; KWTm2, which employs a data-driven approach by integrating delta and delta-delta coefficients along with optimized spectral features, including increased Mel-filter bank resolution and refined frequency boundaries to better capture Turkish phonetic characteristics; and KWTm3, a hybrid model combining both parametric and data-driven enhancements. The study also investigates the efficacy of test-time adaptation techniques to mitigate performance degradation in noisy environments. The experimental findings indicate that the proposed models provide improved accuracy in most noisy conditions and reveal that well-tuned hyperparameters may offer benefits comparable to feature engineering.