Xiaobing Li

Jiafeng Liu, Yuanliang Dong, Zehua Cheng et al.

In this work, we propose a permutation invariant language model, SymphonyNet, as a solution for symbolic symphony music generation. We propose a novel Multi-track Multi-instrument Repeatable (MMR) representation for symphonic music and model the music sequence using a Transformer-based auto-regressive language model with specific 3-D positional embedding. To overcome length overflow when modeling extra-long symphony tokens, we also propose a modified Byte Pair Encoding algorithm (Music BPE) for music tokens and introduce a novel linear transformer decoder architecture as a backbone. Meanwhile, we train the decoder to learn automatic orchestration as a joint task by masking instrument information from the input. We also introduce a large-scale symbolic symphony dataset for the advance of symphony generation research. Empirical results show that the proposed approach can generate coherent, novel, complex and harmonious symphony as a pioneer solution for multi-track multi-instrument symbolic music generation.

Yongsheng Feng, Yuetonghui Xu, Jiehui Luo et al.

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.

Xinran Zhang, Maosong Sun, Jiafeng Liu et al.

Poetry generation has been a difficult task in natural language processing. Unlike plain neural text generation tasks, poetry has a high requirement for novelty, since an easily-understood sentence with too many high frequency words might not be considered as poetic, while adequately ambiguous sentences with low frequency words can possibly be novel and creative. Inspired by this, we present Lingxi, a diversity-aware Chinese modern poetry generation system. We propose nucleus sampling with randomized head (NS-RH) algorithm, which randomizes the high frequency part ("head") of the predicted distribution, in order to emphasize on the "comparatively low frequency" words. The proposed algorithm can significantly increase the novelty of generated poetry compared with traditional sampling methods. The permutation of distribution is controllable by tuning the filtering parameter that determines the "head" to permutate, achieving diversity-aware sampling. We find that even when a large portion of filtered vocabulary is randomized, it can actually generate fluent poetry but with notably higher novelty. We also propose a semantic-similarity-based rejection sampling algorithm, which creates longer and more informative context on the basis of the short input poetry title while maintaining high semantic similarity to the title, alleviating the off-topic issue.

Xinran Zhang, Maosong Sun, Jiafeng Liu et al.

In natural language processing (NLP), the semantic similarity task requires large-scale, high-quality human-annotated labels for fine-tuning or evaluation. By contrast, in cases of music similarity, such labels are expensive to collect and largely dependent on the annotator's artistic preferences. Recent research has demonstrated that embedding calibration technique can greatly increase semantic similarity performance of the pre-trained language model without fine-tuning. However, it is yet unknown which calibration method is the best and how much performance improvement can be achieved. To address these issues, we propose using composer information to construct labels for automatically evaluating music similarity. Under this paradigm, we discover the optimal combination of embedding calibration which achieves superior metrics than the baseline methods.

Xinran Zhang, Maosong Sun, Jiafeng Liu et al.

The neural text generation suffers from the text degeneration issue such as repetition. Traditional stochastic sampling methods only focus on truncating the unreliable "tail" of the distribution, and do not address the "head" part, which we show might contain tedious or even repetitive candidates with high probability that lead to repetition loops. They also do not consider the issue that human text does not always favor high-probability words. Inspired by these, in this work we propose a heuristic sampling method. We propose to use interquartile range of the predicted distribution to determine the "head" part, then permutate and rescale the "head" with inverse probability. This aims at decreasing the probability for the tedious and possibly repetitive candidates with higher probability, and increasing the probability for the rational but more surprising candidates with lower probability. The proposed algorithm provides a reasonable permutation on the predicted distribution which enhances diversity without compromising rationality of the distribution. We use pre-trained language model to compare our algorithm with traditional methods. Results show that our algorithm can effectively increase the diversity of generated samples while achieving close resemblance to human text.