iSPL Lab • Kyungpook National University

Personalized Speech Enhancement: TS-PSE Model

Joint suppression of competing speakers and ambient noise environments

The following are Korean samples. A car conversation recorded during one of conference trips to Jeju Island.

Results & Audio Samples

Sample 1

Noisy Mixture

Target Speaker Reference

Model Output (TS-PSE)

Model Residue (TS-PSE)

Sample 2

Noisy Mixture

Target Speaker Reference

Model Output (TS-PSE)

Model Residue (TS-PSE)

The following are English samples

Results & Audio Samples

Sample 1

Noisy Mixture

Target Speaker Reference

Model Output (TS-PSE)

Sample 2

Noisy Mixture

Target Speaker Reference

Model Output (TS-PSE)

Ground Truth

Sample 3

Noisy Mixture

Target Speaker Reference

Model Output (TS-PSE)

Ground Truth

Sample 4

Noisy Mixture

Target Speaker Reference

Model Output (TS-PSE)

Sample 5

Noisy Mixture

Target Speaker Reference

Model Output (TS-PSE)

Ground Truth

Sample 6

Noisy Mixture

Target Speaker Reference

Model Output (TS-PSE)

Ground Truth

Sample 7

Noisy Mixture

Target Speaker Reference

Model Output (TS-PSE)

Ground Truth

TINYPSE DEMO SAMPLES [AAAI-27]

Sample 1

Noisy Mixture

Target Speaker Reference

Model Output (TinyPSE)

Sample 2

Noisy Mixture

Target Speaker Reference

Model Output (TinyPSE)

Sample 3

Noisy Mixture

Target Speaker Reference

Model Output (TinyPSE)

Sample 4

Noisy Mixture

Target Speaker Reference

Model Output (TinyPSE)

Acknowledgements

We express our sincere gratitude to the following individuals and organizations for their invaluable contributions, feedback, and resource access supporting this research pipeline:

  • Prof. Han Dong Seog For providing continuous expert feedback, resources, and structural guidance throughout this project's lifecycles.
  • Kyungpook National University For offering advanced computational lab clusters and systems infrastructure necessary for handling extensive training configurations.
  • Open-Source Audio Corpora High-quality foundational benchmarking sets provided via Librispeech, VCTK, DEMAND, and various operational vehicle sound banks.

We also extend our appreciation to the broader speech, audio, and signal processing communities for pushing ahead modern architectures in interactive AI.