SpeechWorld at VLSP 2026 SQA: Simulation-Guided DSP–SSL Hybrids for Cross-Lingual Speech Quality Diagnosis

Presented in VLSP 2026, 2026

Authors: Tri-Nhan Do, Long-Hai Tran, Minh-Toan Tran

🏆 First Place in the VLSP 2026 shared task on Speech Quality Assessment (Team SpeechWorld).

The VLSP 2026 Speech Quality Assessment task asks for a quality score and the time-localised detection of noise, distortion and unnatural pauses. We propose a DSP–SSL hybrid that concatenates ten signal-processing frame features with a multilingual self-supervised encoder and is trained on QualiSpeech plus a Vietnamese degradation simulator. An ensemble of four w2v-BERT 2.0 hybrids and a WavLM-based model reaches a final score of 0.617 on the private test. The released labels reveal two shifts: a domain shift in language and fault characteristics, with short and often silent distortion events, and a shift in tag distribution, with unnatural pauses in only 2.6% of the private files against 34–44% in QualiSpeech.

Slides

Paper

Competition Award Certificate

VLSP Competition Award – First Rank, presented at the 12th Workshop on Vietnamese Language and Speech Processing (VLSP 2026), Ho Chi Minh City, Vietnam, November 15, 2026.

VLSP 2026 First Rank certificate – Speech Quality Assessment

At the VLSP 2026 Workshop

VLSP 2026 workshop group photo, Ho Chi Minh City University of Technology