pervasiveaidataresearchlab2025/VN-SpeechMix_Dataset
Viewer • Updated • 26k • 22
AI & ML interests
None defined yet.
Recent Activity
View all activity
Organization Card
Pervasive AI and Data Research Lab
PADRL · Speech & Audio AI
Understanding every voice, even when conversations overlap.
Github · Hugging Face · Contact
We study how AI can recover and process individual voices in overlapping conversations. Our current work connects Vietnamese speech data, neural speech separation, and conversation-processing prototypes, with an emphasis on regional pronunciation differences and reproducible experiments.
Research directions
- Vietnamese speech and dialect diversity: benchmarks spanning Northern, Central, and Southern speech.
- Speech separation: recovering individual speakers from a single audio channel using learned speaker conditioning, temporal modeling, and progressive waveform reconstruction.
- Reproducible audio research: documented data preparation, mixture metadata, reconstruction tools, and controlled model evaluation.
- Speech for conversational applications: exploring separation as a front end for transcription and summarization in meetings, conferencing, and call centers.
Featured research
VN-SpeechMix · Vietnamese speech mixture benchmark
Accepted at ICTA 2026. A benchmark built from the ViMD corpus for studying two-speaker separation across Vietnamese regional dialects.
| Benchmark at a glance | Coverage |
|---|---|
| Mixtures | 26,000 fully overlapped, two-speaker mixtures |
| Duration and audio | Approximately 113 hours, 16 kHz mono |
| Source diversity | 12,048 speakers and 17,401 source utterances used in the mixtures |
| Dialect coverage | 6 balanced pairings across North, Central, and South |
| Train / validation / test | 18,000 / 3,000 / 5,000 mixtures |
The construction pipeline combines source-quality filtering, loudness normalization, dialect-aware pairing, and clipping control. Persisted mixture metadata supports deterministic reconstruction from prepared ViMD-derived recordings.
Resources: Dataset and access conditions · Reconstruction toolkit · Citation record
The Hugging Face release provides gated metadata and scripts. Researchers obtain ViMD separately and follow the terms of both resources when reconstructing audio. The current benchmark uses clean, fully overlapped two-speaker mixtures; real recordings are a focus of further evaluation.
Research community
Our VN-SpeechMix team participated in Saigon AI Hub's first Applied AI Lab cohort, within the programme developed with Google AI Futures Fund. We shared our work at Applied AI Lab Demo Day on 8 October 2026. VNG's official Demo Day report lists VN-SpeechMix among the cohort projects and confirms its ICTA 2026 acceptance.
Follow the programme through Saigon AI Hub on LinkedIn and Facebook.
Team Members
Our team includes:
- Ha Minh Tan · ORCID
- Nguyen Van Hung · Email
- Duong Quoc Nhut · Github
- Nguyen Thi Ngoc Lan · GitHub
Collaborate with us
We welcome collaboration on Vietnamese speech benchmarks, separation methods, ASR evaluation, and representative meeting or call-center recordings. Sustained GPU access and joint downstream experiments can help us extend this work.
models 0
None public yet