- Updated: July 15, 2026
- 2 min read
PS4: Proxy‑Supervised Joint Training for Real Target Speaker Extraction
PS4: Proxy‑Supervised Joint Training for Real Target Speaker Extraction
We are excited to present PS4, a novel proxy‑supervised training framework that pushes the limits of real‑target speaker extraction (TSE) in conversational mixtures. Built on a massive corpus of 71,771 training samples drawn from four public datasets, PS4 covers both Chinese and English scenarios, providing overlapping speech mixtures, per‑speaker enrollment audio, transcripts, and frame‑level voice‑activity labels.
The core of PS4 is a joint training strategy that fine‑tunes a BSRNN‑based TSE model using four complementary, differentiable objectives:
- ASR cross‑entropy loss
- Speaker similarity loss
- Frame‑level voice activity detection loss
- Perceptual audio quality loss
Starting from a publicly available pre‑trained checkpoint, only the BSRNN separator is updated, ensuring efficient adaptation while preserving the strengths of the original model.
On the REAL‑T challenge leaderboard, PS4 achieved a remarkable 2nd place overall, securing the best speaker‑similarity and timing‑F1 scores among all submissions.
Key highlights of PS4:
- Large‑scale, multilingual corpus – 71,771 samples covering diverse acoustic conditions.
- Proxy‑supervised learning – eliminates the need for clean target speech, using readily available proxy labels.
- Joint multi‑objective optimization – balances transcription accuracy, speaker identity preservation, voice activity detection, and perceptual quality.
- Open‑source friendly – built on publicly available datasets and pretrained models.
For a deeper dive into the methodology, implementation details, and performance analysis, read the full paper on arXiv. Stay tuned to ubos.tech for future updates, code releases, and related research.
Author: UBOS Research Team
Andrii Bidochko
CTO UBOS
Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.