# TISDiSS: A Training-Time and Inference-Time Scalable Framework for Discriminative Source Separation

arXiv: https://arxiv.org/abs/2509.15666 · HTML: https://mlsanity.com/paper/2509.15666/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 21 of its 23 references that are in the map.
- First posted on arXiv 19 Sep 2025.

## Builds on (most cited first, 21 of 21)

- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [Deep clustering: Discriminative embeddings for segmentation and separation](https://mlsanity.com/paper/1508.04306/) (2015)
- [Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation](https://mlsanity.com/paper/2407.05361/) (2024)
- [LibriMix: An Open-Source Dataset for Generalizable Speech Separation](https://mlsanity.com/paper/2005.11262/) (2020)
- [Music Source Separation with Band-split RNN](https://mlsanity.com/paper/2209.15174/) (2022)
- [Llasa: Scaling Train-Time and Inference-Time Compute for Llama-based Speech Synthesis](https://mlsanity.com/paper/2502.04128/) (2025)
- [TF-GridNet: Integrating Full- and Sub-Band Modeling for Speech Separation](https://mlsanity.com/paper/2211.12433/) (2022)
- [WHAMR!: Noisy and Reverberant Single-Channel Speech Separation](https://mlsanity.com/paper/1910.10279/) (2019)
- [Music Source Separation with Band-Split RoPE Transformer](https://mlsanity.com/paper/2309.02612/) (2023)
- [ESPnet-se: end-to-end speech enhancement and separation toolkit designed for asr integration](https://mlsanity.com/paper/2011.03706/) (2020)
- [Seed-Music: A Unified Framework for High Quality and Controlled Music Generation](https://mlsanity.com/paper/2409.09214/) (2024)
- [LLaSE-G1: Incentivizing Generalization Capability for LLaMA-based Speech Enhancement](https://mlsanity.com/paper/2503.00493/) (2025)
- [Mel-Band RoFormer for Music Source Separation](https://mlsanity.com/paper/2310.01809/) (2023)
- [AutoPrep: An Automatic Preprocessing Framework for In-the-Wild Speech Data](https://mlsanity.com/paper/2309.13905/) (2023)
- [Separate and Reconstruct: Asymmetric Encoder-Decoder for Speech Separation](https://mlsanity.com/paper/2406.05983/) (2024)
- [TF-Locoformer: Transformer with Local Modeling by Convolution for Speech Separation and Enhancement](https://mlsanity.com/paper/2408.03440/) (2024)
- [TIGER: Time-frequency Interleaved Gain Extraction and Reconstruction for Efficient Speech Separation](https://mlsanity.com/paper/2410.01469/) (2024)
- [Tiny-Sepformer: A Tiny Time-Domain Transformer Network for Speech Separation](https://mlsanity.com/paper/2206.13689/) (2022)
- [Boosting Unknown-number Speaker Separation with Transformer Decoder-based Attractor](https://mlsanity.com/paper/2401.12473/) (2024)
- [Papez: Resource-Efficient Speech Separation with Auditory Working Memory](https://mlsanity.com/paper/2407.00888/) (2024)
- [FLASepformer: Efficient Speech Separation with Gated Focused Linear Attention Transformer](https://mlsanity.com/paper/2508.19528/) (2025)

## Abstract (arXiv)

Source separation is a fundamental task in speech, music, and audio processing, and it also provides cleaner and larger data for training generative models. However, improving separation performance in practice often depends on increasingly large networks, inflating training and deployment costs. Motivated by recent advances in inference-time scaling for generative modeling, we propose Training-Time and Inference-Time Scalable Discriminative Source Separation (TISDiSS), a unified framework that integrates early-split multi-loss supervision, shared-parameter design, and dynamic inference repetitions. TISDiSS enables flexible speed-performance trade-offs by adjusting inference depth without retraining additional models. We further provide systematic analyses of architectural and training choices and show that training with more inference repetitions improves shallow-inference performance, benefiting low-latency applications. Experiments on standard speech separation benchmarks demonstrate state-of-the-art performance with a reduced parameter count, establishing TISDiSS as a scalable and practical framework for adaptive source separation. Code is available at https://github.com/WingSingFung/TISDiSS.