VLSP 2026: Two shared tasks, three first places

Published:

On November 15, 2026, I attended the 12th International Workshop on Vietnamese Language and Speech Processing (VLSP 2026) at Ho Chi Minh City University of Technology (VNUHCM). Our independent team, SpeechWorld, entered two speech shared tasks and won first place in both:

  • Vietnamese Non-Verbal Text-to-Speech for Conversational Synthesis: 1st place in Track A (seen speakers) and 1st place in Track B (zero-shot voice cloning). Our system fine-tuned a multilingual masked generative TTS model with a human-in-the-loop pipeline. Listeners chose every training example, and their preference pairs drove a DPO stage that we adapted to masked token prediction. (paper, audio samples and results)
  • Speech Quality Assessment: 1st place. We used a DSP–SSL hybrid that predicts a quality score and localises noise, distortion and unnatural pauses in time, trained with a Vietnamese degradation simulator. (paper and results)

Thanks to my teammates Viet-Lam Phung, Minh-Khang Le, Hoang-Minh Tran, Thanh-Hau Luong, Long-Hai Tran and Minh-Toan Tran, and to the VLSP organizers.

Presenting the SpeechWorld ViNV-TTS system
Presenting our non-verbal TTS system
Questions from the audience
Q&A session
ViNV-TTS award ceremony, top three teams per track
ViNV-TTS awards: SpeechWorld first in Track A and Track B
Receiving the award certificate
Receiving the award certificate
Poster session
Poster session
Posters along the hallway
Posters along the hallway
Tea break
Tea break
VLSP 2026 group photo
VLSP 2026 group photo

Award Certificates

First Rank certificate, ViNV-TTS Track A First Rank certificate, ViNV-TTS Track B First Rank certificate, Speech Quality Assessment