End-to-end generation of musical audio using deep learning techniques has seen an explosion of activity recently. However, most models concentrate on generating fully mixed music in response to abstract conditioning information. In this work, we present an alternative paradigm for producing music generation models that can listen and respond to musical context. We describe how such a model can be constructed using a non-autoregressive, transformer-based model architecture and present a number of novel architectural and sampling improvements. We train the described architecture on both an open-source and a proprietary dataset. We evaluate the produced models using standard quality metrics and a new approach based on music information retrieval descriptors. The resulting model reaches the audio quality of state-of-the-art text-conditioned models, as well as exhibiting strong musical coherence with its context.
Recently, multi-band spectrogram-based approaches such as Band-Split RNN (BSRNN) have demonstrated promising results for music source separation. In our recent work, we introduce the BS-RoFormer model which inherits the idea of band-split scheme in BSRNN at the front-end, and then uses the hierarchical Transformer with Rotary Position Embedding (RoPE) to model the inner-band and inter-band sequences for multi-band mask estimation. This model has achieved state-of-the-art performance, but the band-split scheme is defined empirically, without analytic supports from the literature. In this paper, we propose Mel-RoFormer, which adopts the Mel-band scheme that maps the frequency bins into overlapped subbands according to the mel scale. In contract, the band-split mapping in BSRNN and BS-RoFormer is non-overlapping and designed based on heuristics. Using the MUSDB18HQ dataset for experiments, we demonstrate that Mel-RoFormer outperforms BS-RoFormer in the separation tasks of vocals, drums, and other stems.
In the era of data-driven Music Information Retrieval (MIR), the scarcity of labeled data has been one of the major concerns to the success of an MIR task. In this work, we leverage the semi-supervised teacher-student training approach to improve MIR tasks. For training, we scale up the unlabeled music data to 240k hours, which is much larger than any public MIR datasets. We iteratively create and refine the pseudo-labels in the noisy teacher-student training process. Knowledge expansion is also explored to iteratively scale up the model sizes from as small as less than 3M to almost 100M parameters. We study the performance correlation between data size and model size in the experiments. By scaling up both model size and training data, our models achieve state-of-the-art results on several MIR tasks compared to models that are either trained in a supervised manner or based on a self-supervised pretrained model. To our knowledge, this is the first attempt to study the effects of scaling up both model and training data for a variety of MIR tasks.
Music source separation (MSS) aims to separate a music recording into multiple musically distinct stems, such as vocals, bass, drums, and more. Recently, deep learning approaches such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been used, but the improvement is still limited. In this paper, we propose a novel frequency-domain approach based on a Band-Split RoPE Transformer (called BS-RoFormer). BS-RoFormer relies on a band-split module to project the input complex spectrogram into subband-level representations, and then arranges a stack of hierarchical Transformers to model the inner-band as well as inter-band sequences for multi-band mask estimation. To facilitate training the model for MSS, we propose to use the Rotary Position Embedding (RoPE). The BS-RoFormer system trained on MUSDB18HQ and 500 extra songs ranked the first place in the MSS track of Sound Demixing Challenge (SDX23). Benchmarking a smaller version of BS-RoFormer on MUSDB18HQ, we achieve state-of-the-art result without extra training data, with 9.80 dB of average SDR.
Multitrack music transcription aims to transcribe a music audio input into the musical notes of multiple instruments simultaneously. It is a very challenging task that typically requires a more complex model to achieve satisfactory result. In addition, prior works mostly focus on transcriptions of regular instruments, however, neglecting vocals, which are usually the most important signal source if present in a piece of music. In this paper, we propose a novel deep neural network architecture, Perceiver TF, to model the time-frequency representation of audio input for multitrack transcription. Perceiver TF augments the Perceiver architecture by introducing a hierarchical expansion with an additional Transformer layer to model temporal coherence. Accordingly, our model inherits the benefits of Perceiver that posses better scalability, allowing it to well handle transcriptions of many instruments in a single model. In experiments, we train a Perceiver TF to model 12 instrument classes as well as vocal in a multi-task learning manner. Our result demonstrates that the proposed system outperforms the state-of-the-art counterparts (e.g., MT3 and SpecTNT) on various public datasets.
Singing voice beat and downbeat tracking posses several applications in automatic music production, analysis and manipulation. Among them, some require real-time processing, such as live performance processing and auto-accompaniment for singing inputs. This task is challenging owing to the non-trivial rhythmic and harmonic patterns in singing signals. For real-time processing, it introduces further constraints such as inaccessibility to future data and the impossibility to correct the previous results that are inconsistent with the latter ones. In this paper, we introduce the first system that tracks the beats and downbeats of singing voices in real-time. Specifically, we propose a novel dynamic particle filtering approach that incorporates offline historical data to correct the online inference by using a variable number of particles. We evaluate the performance on two datasets: GTZAN with the separated vocal tracks, and an in-house dataset with the original vocal stems. Experimental result demonstrates that our proposed approach outperforms the baseline by 3-5%.
In this paper, we introduce Jointist, an instrument-aware multi-instrument framework that is capable of transcribing, recognizing, and separating multiple musical instruments from an audio clip. Jointist consists of an instrument recognition module that conditions the other two modules: a transcription module that outputs instrument-specific piano rolls, and a source separation module that utilizes instrument information and transcription results. The joint training of the transcription and source separation modules serves to improve the performance of both tasks. The instrument module is optional and can be directly controlled by human users. This makes Jointist a flexible user-controllable framework. Our challenging problem formulation makes the model highly useful in the real world given that modern popular music typically consists of multiple instruments. Its novelty, however, necessitates a new perspective on how to evaluate such a model. In our experiments, we assess the proposed model from various aspects, providing a new evaluation perspective for multi-instrument transcription. Our subjective listening study shows that Jointist achieves state-of-the-art performance on popular music, outperforming existing multi-instrument transcription models such as MT3. We conducted experiments on several downstream tasks and found that the proposed method improved transcription by more than 1 percentage points (ppt.), source separation by 5 SDR, downbeat detection by 1.8 ppt., chord recognition by 1.4 ppt., and key estimation by 1.4 ppt., when utilizing transcription results obtained from Jointist. Demo available at \url{https://jointist.github.io/Demo}.
Creating a pop song melody according to pre-written lyrics is a typical practice for composers. A computational model of how lyrics are set as melodies is important for automatic composition systems, but an end-to-end lyric-to-melody model would require enormous amounts of paired training data. To mitigate the data constraints, we adopt a two-stage approach, dividing the task into lyric-to-rhythm and rhythm-to-melody modules. However, the lyric-to-rhythm task is still challenging due to its multimodality. In this paper, we propose a novel lyric-to-rhythm framework that includes part-of-speech tags to achieve better text setting, and a Transformer architecture designed to model long-term syllable-to-note associations. For the rhythm-to-melody task, we adapt a proven chord-conditioned melody Transformer, which has achieved state-of-the-art results. Experiments for Chinese lyric-to-melody generation show that the proposed framework is able to model key characteristics of rhythm and pitch distributions in the dataset, and in a subjective evaluation, the melodies generated by our system were rated as similar to or better than those of a state-of-the-art alternative.
Music structure analysis (MSA) systems aim to segment a song recording into non-overlapping sections with useful labels. Previous MSA systems typically predict abstract labels in a post-processing step and require the full context of the song. By contrast, we recently proposed a supervised framework, called "Music Structural Function Analysis" (MuSFA), that models and predicts meaningful labels like 'verse' and 'chorus' directly from audio, without requiring the full context of a song. However, the performance of this system depends on the amount and quality of training data. In this paper, we propose to repurpose a public dataset, HookTheory Lead Sheet Dataset (HLSD), to improve the performance. HLSD contains over 18K excerpts of music sections originally collected for studying automatic melody harmonization. We treat each excerpt as a partially labeled song and provide a label mapping, so that HLSD can be used together with other public datasets, such as SALAMI, RWC, and Isophonics. In cross-dataset evaluations, we find that including HLSD in training can improve state-of-the-art boundary detection and section labeling scores by ~3% and ~1% respectively.