Music is an integral part of human culture, embodying human intelligence and creativity, of which songs compose an essential part. While various aspects of song generation have been explored by previous works, such as singing voice, vocal composition and instrumental arrangement, etc., generating songs with both vocals and accompaniment given lyrics remains a significant challenge, hindering the application of music generation models in the real world. In this light, we propose SongCreator, a song-generation system designed to tackle this challenge. The model features two novel designs: a meticulously designed dual-sequence language model (DSLM) to capture the information of vocals and accompaniment for song generation, and a series of attention mask strategies for DSLM, which allows our model to understand, generate and edit songs, making it suitable for various songrelated generation tasks by utilizing specific attention masks. Extensive experiments demonstrate the effectiveness of SongCreator by achieving state-of-the-art or competitive performances on all eight tasks. Notably, it surpasses previous works by a large margin in lyrics-to-song and lyrics-to-vocals. Additionally, it is able to independently control the acoustic conditions of the vocals and accompaniment in the generated song through different audio prompts, exhibiting its potential applicability. Our samples are available at https://thuhcsi.github.io/SongCreator/.

Recent progress in music generation has been remarkably advanced by the state-of-the-art MusicLM, which comprises a hierarchy of three LMs, respectively, for semantic, coarse acoustic, and fine acoustic modelings. Yet, sampling with the MusicLM requires processing through these LMs one by one to obtain the fine-grained acoustic tokens, making it computationally expensive and prohibitive for a real-time generation. Efficient music generation with a quality on par with MusicLM remains a significant challenge. In this paper, we present **M**e**L**o**D**y (**M** for music; **L** for LM; **D** for diffusion), an LM-guided diffusion model that generates music audios of state-of-the-art quality meanwhile reducing 95.7\% to 99.6\% forward passes in MusicLM, respectively, for sampling 10s to 30s music. MeLoDy inherits the highest-level LM from MusicLM for semantic modeling, and applies a novel dual-path diffusion (DPD) model and an audio VAE-GAN to efficiently decode the conditioning semantic tokens into waveform. DPD is proposed to simultaneously model the coarse and fine acoustics by incorporating the semantic information into segments of latents effectively via cross-attention at each denoising step. Our experimental results suggest the superiority of MeLoDy, not only in its practical advantages on sampling speed and infinitely continuable generation, but also in its state-of-the-art musicality, audio quality, and text correlation. Our samples are available at https://Efficient-MeLoDy.github.io/.

Denoising diffusion probabilistic models (DDPMs) have recently achieved leading performances in many generative tasks. However, the inherited iterative sampling process costs hindered their applications to speech synthesis. This paper proposes FastDiff, a fast conditional diffusion model for high-quality speech synthesis. FastDiff employs a stack of time-aware location-variable convolutions of diverse receptive field patterns to efficiently model long-term time dependencies with adaptive conditions. A noise schedule predictor is also adopted to reduce the sampling steps without sacrificing the generation quality. Based on FastDiff, we design an end-to-end text-to-speech synthesizer, FastDiff-TTS, which generates high-fidelity speech waveforms without any intermediate feature (e.g., Mel-spectrogram). Our evaluation of FastDiff demonstrates the state-of-the-art results with higher-quality (MOS 4.28) speech samples. Also, FastDiff enables a sampling speed of 58x faster than real-time on a V100 GPU, making diffusion models practically applicable to speech synthesis deployment for the first time. We further show that FastDiff generalized well to the mel-spectrogram inversion of unseen speakers, and FastDiff-TTS outperformed other competing methods in end-to-end text-to-speech synthesis. Audio samples are available at https://FastDiff.github.io/.

We study the cocktail party problem and propose a novel attention network called Tune-In, abbreviated for training under negative environments with interference. It firstly learns two separate spaces of speaker-knowledge and speech-stimuli based on a shared feature space, where a new block structure is designed as the building block for all spaces, and then cooperatively solves different tasks. Between the two spaces, information is cast towards each other via a novel cross- and dual-attention mechanism, mimicking the bottom-up and top-down processes of a human's cocktail party effect. It turns out that substantially discriminative and generalizable speaker representations can be learnt in severely interfered conditions via our self-supervised training. The experimental results verify this seeming paradox. The learnt speaker embedding has superior discriminative power than a standard speaker verification method; meanwhile, Tune-In achieves remarkably better speech separation performances in terms of SI-SNRi and SDRi consistently in all test modes, and especially at lower memory and computational consumption, than state-of-the-art benchmark systems.

One of the leading single-channel speech separation (SS) models is based on a TasNet with a dual-path segmentation technique, where the size of each segment remains unchanged throughout all layers. In contrast, our key finding is that multi-granularity features are essential for enhancing contextual modeling and computational efficiency. We introduce a self-attentive network with a novel sandglass-shape, namely Sandglasset, which advances the state-of-the-art (SOTA) SS performance at significantly smaller model size and computational cost. Forward along each block inside Sandglasset, the temporal granularity of the features gradually becomes coarser until reaching half of the network blocks, and then successively turns finer towards the raw signal level. We also unfold that residual connections between features with the same granularity are critical for preserving information after passing through the bottleneck layer. Experiments show our Sandglasset with only 2.3M parameters has achieved the best results on two benchmark SS datasets – WSJ0-2mix and WSJ0-3mix, where the SI-SNRi scores have been improved by absolute 0.6 dB and 2.4 dB, respectively, comparing to the prior SOTA results.

To extract robust deep representations from long sequential modeling of speech data, we propose a self-supervised learning approach, namely Contrastive Separative Coding (CSC). Our key finding is to learn such representations by separating the target signal from contrastive interfering signals. First, a multi-task separative encoder is built to extract shared separable and discriminative embedding; secondly, we propose a powerful cross-attention mechanism performed over speaker representations across various interfering conditions, allowing the model to focus on and globally aggregate the most critical information to answer the "query" (current bottom-up embedding) while paying less attention to interfering, noisy, or irrelevant parts; lastly, we form a new probabilistic contrastive loss which estimates and maximizes the mutual information between the representations and the global speaker vector. While most prior unsupervised methods have focused on predicting the future, neighboring, or missing samples, we take a different perspective of predicting the interfered samples. Moreover, our contrastive separative loss is free from negative sampling. The experiment demonstrates that our approach can learn useful representations achieving a strong speaker verification performance in adverse conditions.

End-to-end speech recognition generally uses hand-engineered acoustic features as input and excludes the feature extraction module from its joint optimization.To extract learnable and adaptive features and mitigate information loss, we propose a new encoder that adopts globally attentive locally recurrent (GALR) networks and directly takes raw waveform as input.We observe improved ASR performance and robustness by applying GALR on different window lengths to aggregate fine-grain temporal information into multi-scale acoustic features.Experiments are conducted on a benchmark dataset AISHELL-2 and two large-scale Mandarin speech corpus of 5, 000 hours and 21, 000 hours.With faster speed and comparable model size, our proposed multi-scale GALR waveform encoder achieved consistent character error rate reductions (CERRs) from 7.9% to 28.1% relative over strong baselines, including Conformer and TDNN-Conformer.In particular, our approach demonstrated notable robustness than the traditional handcrafted features and outperformed the baseline MFCC-based TDNN-Conformer model by a 15.2% CERR on a music-mixed real-world speech test set.

Recent research on the time-domain audio separation networks (TasNets) has brought great success to speech separation. Nevertheless, conventional TasNets struggle to satisfy the memory and latency constraints in industrial applications. In this regard, we design a low-cost high-performance architecture, namely, globally attentive locally recurrent (GALR) network. Alike the dual-path RNN (DPRNN), we first split a feature sequence into 2D segments and then process the sequence along both the intra- and inter-segment dimensions. Our main innovation lies in that, on top of features recurrently processed along the inter-segment dimensions, GALR applies a self-attention mechanism to the sequence along the inter-segment dimension, which aggregates context-aware information and also enables parallelization. Our experiments suggest that GALR is a notably more effective network than the prior work. On one hand, with only 1.5M parameters, it has achieved comparable separation performance at a much lower cost with 36.1% less runtime memory and 49.4% fewer computational operations, relative to the DPRNN. On the other hand, in a comparable model size with DPRNN, GALR has consistently outperformed DPRNN in three datasets, in particular, with a substantial margin of 2.4dB absolute improvement of SI-SNRi in the benchmark WSJ0-2mix task.

Deep-learning based speech separation models confront poor generalization problem that even the state-of-the-art models could abruptly fail when evaluating them in mismatch conditions. To address this problem, we propose an easy-to-implement yet effective consistency based semi-supervised learning (SSL) approach, namely Mixup-Breakdown training (MBT). It learns a teacher model to "breakdown" unlabeled inputs, and the estimated separations are interpolated to produce more useful pseudo "mixup" input-output pairs, on which the consistency regularization could apply for learning a student model. In our experiment, we evaluate MBT under various conditions with ascending degrees of mismatch, including unseen interfering speech, noise, and music, and compare MBT’s generalization capability against state-of-the-art supervised learning and SSL approaches. The result indicates that MBT significantly outperforms several strong baselines with up to 13.77% relative SI-SNRi improvement. Moreover, MBT only adds negligible computational overhead to standard training schemes.

The hidden activation functions inside deep neural networks (DNNs) play a vital role in learning high level discriminative features and controlling the information flows to track longer history. However, the fixed model parameters used in standard DNNs can lead to over-fitting and poor generalization when given limited training data. Furthermore, the precise forms of activations used in DNNs are often manually set at a global level for all hidden nodes, thus lacking an automatic selection method. In order to address these issues, Bayesian neural networks (BNNs) acoustic models are proposed in this paper to explicitly model the uncertainty associated with DNN parameters. Gaussian Process (GP) activations based DNN and LSTM acoustic models are also used in this paper to allow the optimal forms of hidden activations to be stochastically learned for individual hidden nodes. An efficient variational inference based training algorithm is derived for BNN, GPNN and GPLSTM systems. Experiments were conducted on a LVCSR system trained on a 75 hour subset of Switchboard I data. The best BNN and GPNN systems outperformed both the baseline DNN systems constructed using fixed form activations and their combination via frame level joint decoding by 1% absolute in word error rate.

Recurrent neural network language models (RNNLMs) have shown superior performance across a range of speech recognition tasks. At the heart of all RNNLMs, the activation functions play a vital role to control the information flows and tracking longer history contexts that are useful for predicting the following words. Long short-term memory (LSTM) units are well known for such ability and thus widely used in current RNNLMs. However, the deterministic parameter estimates in LSTM RNNLMs are prone to over-fitting and poor generalization when given limited training data. Furthermore, the precise forms of activations in LSTM have been largely empirically set for all cells at a global level. In order to address these issues, this paper introduces Gaussian process (GP) LSTM RNNLMs. In addition to modeling parameter uncertainty under a Bayesian framework, it also allows the optimal forms of gates being automatically learned for individual LSTM cells. Experiments were conducted on three tasks: the Penn Treebank (PTB) corpus, Switchboard conversational telephone speech (SWBD) and the AMI meeting room data. The proposed GP-LSTM RNNLMs consistently outperform the baseline LSTM RNNLMs in terms of both perplexity and word error rate.

Recurrent neural network language models (RNNLMs) have become an increasing popular choice for state-of-the-art speech recognition systems. RNNLMs are normally trained by minimizing the cross entropy (CE) using the stochastic gradient descent (SGD) algorithm. However, the SGD method doesn't consider the correlation between parameters and therefore can lead to unstable and slow convergence in training. Second-order optimization methods provide a possible solution to this issue. However these methods are either computationally heavy or do not have competitive performance. In this paper, a novel optimization method - stochastic natural gradient based on minimum variance assumption (SNGM) is proposed for training RNNLMs. It allows the natural gradient method to operate at a comparable training efficiency to the SGD method. By modifying the gradient according to the local curvature of the KL-divergence between current and updated probabilistic distributions, the proposed SNGM approach is shown to outperform both the SGD and limited memory BFGS methods across three tasks: Penn Treebank, Switchboard conversational speech recognition and AMI meeting room transcription in terms of both perplexity and word error rate.