Haohe Liu

Yichong Leng, Zehua Chen, Junliang Guo et al. · microsoft-research

Binaural audio plays a significant role in constructing immersive augmented and virtual realities. As it is expensive to record binaural audio from the real world, synthesizing them from mono audio has attracted increasing attention. This synthesis process involves not only the basic physical warping of the mono audio, but also room reverberations and head/ear related filtrations, which, however, are difficult to accurately simulate in traditional digital signal processing. In this paper, we formulate the synthesis process from a different perspective by decomposing the binaural audio into a common part that shared by the left and right channels as well as a specific part that differs in each channel. Accordingly, we propose BinauralGrad, a novel two-stage framework equipped with diffusion models to synthesize them respectively. Specifically, in the first stage, the common information of the binaural audio is generated with a single-channel diffusion model conditioned on the mono audio, based on which the binaural audio is generated by a two-channel diffusion model in the second stage. Combining this novel perspective of two-stage synthesis with advanced generative models (i.e., the diffusion models),the proposed BinauralGrad is able to generate accurate and high-fidelity binaural audio samples. Experiment results show that on a benchmark dataset, BinauralGrad outperforms the existing baselines by a large margin in terms of both object and subject evaluation metrics (Wave L2: 0.128 vs. 0.157, MOS: 3.80 vs. 3.61). The generated audio samples (https://speechresearch.github.io/binauralgrad) and code (https://github.com/microsoft/NeuralSpeech/tree/master/BinauralGrad) are available online.

Xubo Liu, Qiuqiang Kong, Yan Zhao et al. · bytedance

Language-queried audio source separation (LASS) is a new paradigm for computational auditory scene analysis (CASA). LASS aims to separate a target sound from an audio mixture given a natural language query, which provides a natural and scalable interface for digital audio applications. Recent works on LASS, despite attaining promising separation performance on specific sources (e.g., musical instruments, limited classes of audio events), are unable to separate audio concepts in the open domain. In this work, we introduce AudioSep, a foundation model for open-domain audio source separation with natural language queries. We train AudioSep on large-scale multimodal datasets and extensively evaluate its capabilities on numerous tasks including audio event separation, musical instrument separation, and speech enhancement. AudioSep demonstrates strong separation performance and impressive zero-shot generalization ability using audio captions or text labels as queries, substantially outperforming previous audio-queried and language-queried sound separation models. For reproducibility of this work, we will release the source code, evaluation benchmark and pre-trained model at: https://github.com/Audio-AGI/AudioSep.

Xu Tan, Jiawei Chen, Haohe Liu et al. · microsoft-research

Text to speech (TTS) has made rapid progress in both academia and industry in recent years. Some questions naturally arise that whether a TTS system can achieve human-level quality, how to define/judge that quality and how to achieve it. In this paper, we answer these questions by first defining the human-level quality based on the statistical significance of subjective measure and introducing appropriate guidelines to judge it, and then developing a TTS system called NaturalSpeech that achieves human-level quality on a benchmark dataset. Specifically, we leverage a variational autoencoder (VAE) for end-to-end text to waveform generation, with several key modules to enhance the capacity of the prior from text and reduce the complexity of the posterior from speech, including phoneme pre-training, differentiable duration modeling, bidirectional prior/posterior modeling, and a memory mechanism in VAE. Experiment evaluations on popular LJSpeech dataset show that our proposed NaturalSpeech achieves -0.01 CMOS (comparative mean opinion score) to human recordings at the sentence level, with Wilcoxon signed rank test at p-level p >> 0.05, which demonstrates no statistically significant difference from human recordings for the first time on this dataset.

Xinhao Mei, Chutong Meng, Haohe Liu et al.

The advancement of audio-language (AL) multimodal learning tasks has been significant in recent years. However, researchers face challenges due to the costly and time-consuming collection process of existing audio-language datasets, which are limited in size. To address this data scarcity issue, we introduce WavCaps, the first large-scale weakly-labelled audio captioning dataset, comprising approximately 400k audio clips with paired captions. We sourced audio clips and their raw descriptions from web sources and a sound event detection dataset. However, the online-harvested raw descriptions are highly noisy and unsuitable for direct use in tasks such as automated audio captioning. To overcome this issue, we propose a three-stage processing pipeline for filtering noisy data and generating high-quality captions, where ChatGPT, a large language model, is leveraged to filter and transform raw descriptions automatically. We conduct a comprehensive analysis of the characteristics of WavCaps dataset and evaluate it on multiple downstream audio-language multimodal learning tasks. The systems trained on WavCaps outperform previous state-of-the-art (SOTA) models by a significant margin. Our aspiration is for the WavCaps dataset we have proposed to facilitate research in audio-language multimodal learning and demonstrate the potential of utilizing ChatGPT to enhance academic research. Our dataset and codes are available at https://github.com/XinhaoMei/WavCaps.

Haohe Liu, Xubo Liu, Xinhao Mei et al.

Few-shot bioacoustic event detection is a task that detects the occurrence time of a novel sound given a few examples. Previous methods employ metric learning to build a latent space with the labeled part of different sound classes, also known as positive events. In this study, we propose a segment-level few-shot learning framework that utilizes both the positive and negative events during model optimization. Training with negative events, which are larger in volume than positive events, can increase the generalization ability of the model. In addition, we use transductive inference on the validation set during training for better adaptation to novel classes. We conduct ablation studies on our proposed method with different setups on input features, training data, and hyper-parameters. Our final system achieves an F-measure of 62.73 on the DCASE 2022 challenge task 5 (DCASE2022-T5) validation set, outperforming the performance of the baseline prototypical network 34.02 by a large margin. Using the proposed method, our submitted system ranks 2nd in DCASE2022-T5. The code of this paper is fully open-sourced at https://github.com/haoheliu/DCASE_2022_Task_5.

Haohe Liu, Zehua Chen, Yi Yuan et al.

Text-to-audio (TTA) system has recently gained attention for its ability to synthesize general audio based on text descriptions. However, previous studies in TTA have limited generation quality with high computational costs. In this study, we propose AudioLDM, a TTA system that is built on a latent space to learn the continuous audio representations from contrastive language-audio pretraining (CLAP) latents. The pretrained CLAP models enable us to train LDMs with audio embedding while providing text embedding as a condition during sampling. By learning the latent representations of audio signals and their compositions without modeling the cross-modal relationship, AudioLDM is advantageous in both generation quality and computational efficiency. Trained on AudioCaps with a single GPU, AudioLDM achieves state-of-the-art TTA performance measured by both objective and subjective metrics (e.g., frechet distance). Moreover, AudioLDM is the first TTA system that enables various text-guided audio manipulations (e.g., style transfer) in a zero-shot fashion. Our implementation and demos are available at https://audioldm.github.io.

Haohe Liu, Yi Yuan, Xubo Liu et al.

Although audio generation shares commonalities across different types of audio, such as speech, music, and sound effects, designing models for each type requires careful consideration of specific objectives and biases that can significantly differ from those of other types. To bring us closer to a unified perspective of audio generation, this paper proposes a framework that utilizes the same learning method for speech, music, and sound effect generation. Our framework introduces a general representation of audio, called "language of audio" (LOA). Any audio can be translated into LOA based on AudioMAE, a self-supervised pre-trained representation learning model. In the generation process, we translate any modalities into LOA by using a GPT-2 model, and we perform self-supervised audio generation learning with a latent diffusion model conditioned on LOA. The proposed framework naturally brings advantages such as in-context learning abilities and reusable self-supervised pretrained AudioMAE and latent diffusion models. Experiments on the major benchmarks of text-to-audio, text-to-music, and text-to-speech demonstrate state-of-the-art or competitive performance against previous approaches. Our code, pretrained model, and demo are available at https://audioldm.github.io/audioldm2.

Haohe Liu, Woosung Choi, Xubo Liu et al.

Speech super-resolution (SR) is a task to increase speech sampling rate by generating high-frequency components. Existing speech SR methods are trained in constrained experimental settings, such as a fixed upsampling ratio. These strong constraints can potentially lead to poor generalization ability in mismatched real-world cases. In this paper, we propose a neural vocoder based speech super-resolution method (NVSR) that can handle a variety of input resolution and upsampling ratios. NVSR consists of a mel-bandwidth extension module, a neural vocoder module, and a post-processing module. Our proposed system achieves state-of-the-art results on the VCTK multi-speaker benchmark. On 44.1 kHz target resolution, NVSR outperforms WSRGlow and Nu-wave by 8% and 37% respectively on log spectral distance and achieves a significantly better perceptual quality. We also demonstrate that prior knowledge in the pre-trained vocoder is crucial for speech SR by performing mel-bandwidth extension with a simple replication-padding method. Samples can be found in https://haoheliu.github.io/nvsr.

Xubo Liu, Haohe Liu, Qiuqiang Kong et al.

In this paper, we introduce the task of language-queried audio source separation (LASS), which aims to separate a target source from an audio mixture based on a natural language query of the target source (e.g., "a man tells a joke followed by people laughing"). A unique challenge in LASS is associated with the complexity of natural language description and its relation with the audio sources. To address this issue, we proposed LASS-Net, an end-to-end neural network that is learned to jointly process acoustic and linguistic information, and separate the target source that is consistent with the language query from an audio mixture. We evaluate the performance of our proposed system with a dataset created from the AudioCaps dataset. Experimental results show that LASS-Net achieves considerable improvements over baseline methods. Furthermore, we observe that LASS-Net achieves promising generalization results when using diverse human-annotated descriptions as queries, indicating its potential use in real-world scenarios. The separated audio samples and source code are available at https://liuxubo717.github.io/LASS-demopage.

Haohe Liu, Ke Chen, Qiao Tian et al.

Audio super-resolution is a fundamental task that predicts high-frequency components for low-resolution audio, enhancing audio quality in digital applications. Previous methods have limitations such as the limited scope of audio types (e.g., music, speech) and specific bandwidth settings they can handle (e.g., 4kHz to 8kHz). In this paper, we introduce a diffusion-based generative model, AudioSR, that is capable of performing robust audio super-resolution on versatile audio types, including sound effects, music, and speech. Specifically, AudioSR can upsample any input audio signal within the bandwidth range of 2kHz to 16kHz to a high-resolution audio signal at 24kHz bandwidth with a sampling rate of 48kHz. Extensive objective evaluation on various audio super-resolution benchmarks demonstrates the strong result achieved by the proposed model. In addition, our subjective evaluation shows that AudioSR can acts as a plug-and-play module to enhance the generation quality of a wide range of audio generative models, including AudioLDM, Fastspeech2, and MusicGen. Our code and demo are available at https://audioldm.github.io/audiosr.

Yi Yuan, Haohe Liu, Xubo Liu et al.

Despite recent progress in text-to-audio (TTA) generation, we show that the state-of-the-art models, such as AudioLDM, trained on datasets with an imbalanced class distribution, such as AudioCaps, are biased in their generation performance. Specifically, they excel in generating common audio classes while underperforming in the rare ones, thus degrading the overall generation performance. We refer to this problem as long-tailed text-to-audio generation. To address this issue, we propose a simple retrieval-augmented approach for TTA models. Specifically, given an input text prompt, we first leverage a Contrastive Language Audio Pretraining (CLAP) model to retrieve relevant text-audio pairs. The features of the retrieved audio-text data are then used as additional conditions to guide the learning of TTA models. We enhance AudioLDM with our proposed approach and denote the resulting augmented system as Re-AudioLDM. On the AudioCaps dataset, Re-AudioLDM achieves a state-of-the-art Frechet Audio Distance (FAD) of 1.37, outperforming the existing approaches by a large margin. Furthermore, we show that Re-AudioLDM can generate realistic audio for complex scenes, rare audio classes, and even unseen audio types, indicating its potential in TTA tasks.

Xubo Liu, Zhongkai Zhu, Haohe Liu et al.

Despite breakthroughs in audio generation models, their capabilities are often confined to domain-specific conditions such as speech transcriptions and audio captions. However, real-world audio creation aims to generate harmonious audio containing various elements such as speech, music, and sound effects with controllable conditions, which is challenging to address using existing audio generation systems. We present WavJourney, a novel framework that leverages Large Language Models (LLMs) to connect various audio models for audio creation. WavJourney allows users to create storytelling audio content with diverse audio elements simply from textual descriptions. Specifically, given a text instruction, WavJourney first prompts LLMs to generate an audio script that serves as a structured semantic representation of audio elements. The audio script is then converted into a computer program, where each line of the program calls a task-specific audio generation model or computational operation function. The computer program is then executed to obtain a compositional and interpretable solution for audio creation. Experimental results suggest that WavJourney is capable of synthesizing realistic audio aligned with textually-described semantic, spatial and temporal conditions, achieving state-of-the-art results on text-to-audio generation benchmarks. Additionally, we introduce a new multi-genre story benchmark. Subjective evaluations demonstrate the potential of WavJourney in crafting engaging storytelling audio content from text. We further demonstrate that WavJourney can facilitate human-machine co-creation in multi-round dialogues. To foster future research, the code and synthesized audio are available at: https://audio-agi.github.io/WavJourney_demopage/.

Xubo Liu, Xinhao Mei, Qiushi Huang et al.

Audio captioning aims at using natural language to describe the content of an audio clip. Existing audio captioning systems are generally based on an encoder-decoder architecture, in which acoustic information is extracted by an audio encoder and then a language decoder is used to generate the captions. Training an audio captioning system often encounters the problem of data scarcity. Transferring knowledge from pre-trained audio models such as Pre-trained Audio Neural Networks (PANNs) have recently emerged as a useful method to mitigate this issue. However, there is less attention on exploiting pre-trained language models for the decoder, compared with the encoder. BERT is a pre-trained language model that has been extensively used in Natural Language Processing (NLP) tasks. Nevertheless, the potential of BERT as the language decoder for audio captioning has not been investigated. In this study, we demonstrate the efficacy of the pre-trained BERT model for audio captioning. Specifically, we apply PANNs as the encoder and initialize the decoder from the public pre-trained BERT models. We conduct an empirical study on the use of these BERT models for the decoder in the audio captioning model. Our models achieve competitive results with the existing audio captioning methods on the AudioCaps dataset.

Ke Chen, Yusong Wu, Haohe Liu et al.

Diffusion models have shown promising results in cross-modal generation tasks, including text-to-image and text-to-audio generation. However, generating music, as a special type of audio, presents unique challenges due to limited availability of music data and sensitive issues related to copyright and plagiarism. In this paper, to tackle these challenges, we first construct a state-of-the-art text-to-music model, MusicLDM, that adapts Stable Diffusion and AudioLDM architectures to the music domain. We achieve this by retraining the contrastive language-audio pretraining model (CLAP) and the Hifi-GAN vocoder, as components of MusicLDM, on a collection of music data samples. Then, to address the limitations of training data and to avoid plagiarism, we leverage a beat tracking model and propose two different mixup strategies for data augmentation: beat-synchronous audio mixup and beat-synchronous latent mixup, which recombine training audio directly or via a latent embeddings space, respectively. Such mixup strategies encourage the model to interpolate between musical training samples and generate new music within the convex hull of the training data, making the generated music more diverse while still staying faithful to the corresponding style. In addition to popular evaluation metrics, we design several new evaluation metrics based on CLAP score to demonstrate that our proposed MusicLDM and beat-synchronous mixup strategies improve both the quality and novelty of generated music, as well as the correspondence between input text and generated music.

Xubo Liu, Qiushi Huang, Xinhao Mei et al.

Audio captioning aims to generate text descriptions of audio clips. In the real world, many objects produce similar sounds. How to accurately recognize ambiguous sounds is a major challenge for audio captioning. In this work, inspired by inherent human multimodal perception, we propose visually-aware audio captioning, which makes use of visual information to help the description of ambiguous sounding objects. Specifically, we introduce an off-the-shelf visual encoder to extract video features and incorporate the visual features into an audio captioning system. Furthermore, to better exploit complementary audio-visual contexts, we propose an audio-visual attention mechanism that adaptively integrates audio and visual context and removes the redundant information in the latent space. Experimental results on AudioCaps, the largest audio captioning dataset, show that our proposed method achieves state-of-the-art results on machine translation metrics.

Xubo Liu, Haohe Liu, Qiuqiang Kong et al.

Recently, there has been increasing interest in building efficient audio neural networks for on-device scenarios. Most existing approaches are designed to reduce the size of audio neural networks using methods such as model pruning. In this work, we show that instead of reducing model size using complex methods, eliminating the temporal redundancy in the input audio features (e.g., mel-spectrogram) could be an effective approach for efficient audio classification. To do so, we proposed a family of simple pooling front-ends (SimPFs) which use simple non-parametric pooling operations to reduce the redundant information within the mel-spectrogram. We perform extensive experiments on four audio classification tasks to evaluate the performance of SimPFs. Experimental results show that SimPFs can achieve a reduction in more than half of the number of floating point operations (FLOPs) for off-the-shelf audio neural networks, with negligible degradation or even some improvements in audio classification performance.

Zehua Chen, Yihan Wu, Yichong Leng et al.

Denoising Diffusion Probabilistic Models (DDPMs) are emerging in text-to-speech (TTS) synthesis because of their strong capability of generating high-fidelity samples. However, their iterative refinement process in high-dimensional data space results in slow inference speed, which restricts their application in real-time systems. Previous works have explored speeding up by minimizing the number of inference steps but at the cost of sample quality. In this work, to improve the inference speed for DDPM-based TTS model while achieving high sample quality, we propose ResGrad, a lightweight diffusion model which learns to refine the output spectrogram of an existing TTS model (e.g., FastSpeech 2) by predicting the residual between the model output and the corresponding ground-truth speech. ResGrad has several advantages: 1) Compare with other acceleration methods for DDPM which need to synthesize speech from scratch, ResGrad reduces the complexity of task by changing the generation target from ground-truth mel-spectrogram to the residual, resulting into a more lightweight model and thus a smaller real-time factor. 2) ResGrad is employed in the inference process of the existing TTS model in a plug-and-play way, without re-training this model. We verify ResGrad on the single-speaker dataset LJSpeech and two more challenging datasets with multiple speakers (LibriTTS) and high sampling rate (VCTK). Experimental results show that in comparison with other speed-up methods of DDPMs: 1) ResGrad achieves better sample quality with the same inference speed measured by real-time factor; 2) with similar speech quality, ResGrad synthesizes speech faster than baseline methods by more than 10 times. Audio samples are available at https://resgrad1.github.io/.

Haohe Liu, Xubo Liu, Qiuqiang Kong et al.

The audio spectrogram is a time-frequency representation that has been widely used for audio classification. One of the key attributes of the audio spectrogram is the temporal resolution, which depends on the hop size used in the Short-Time Fourier Transform (STFT). Previous works generally assume the hop size should be a constant value (e.g., 10 ms). However, a fixed temporal resolution is not always optimal for different types of sound. The temporal resolution affects not only classification accuracy but also computational cost. This paper proposes a novel method, DiffRes, that enables differentiable temporal resolution modeling for audio classification. Given a spectrogram calculated with a fixed hop size, DiffRes merges non-essential time frames while preserving important frames. DiffRes acts as a "drop-in" module between an audio spectrogram and a classifier and can be jointly optimized with the classification task. We evaluate DiffRes on five audio classification tasks, using mel-spectrograms as the acoustic features, followed by off-the-shelf classifier backbones. Compared with previous methods using the fixed temporal resolution, the DiffRes-based method can achieve the equivalent or better classification accuracy with at least 25% computational cost reduction. We further show that DiffRes can improve classification accuracy by increasing the temporal resolution of input acoustic features, without adding to the computational cost.

Yi Yuan, Haohe Liu, Jinhua Liang et al.

Deep neural networks have recently achieved breakthroughs in sound generation. Despite the outstanding sample quality, current sound generation models face issues on small-scale datasets (e.g., overfitting), significantly limiting performance. In this paper, we make the first attempt to investigate the benefits of pre-training on sound generation with AudioLDM, the cutting-edge model for audio generation, as the backbone. Our study demonstrates the advantages of the pre-trained AudioLDM, especially in data-scarcity scenarios. In addition, the baselines and evaluation protocol for sound generation systems are not consistent enough to compare different studies directly. Aiming to facilitate further study on sound generation tasks, we benchmark the sound generation task on various frequently-used datasets. We hope our results on transfer learning and benchmarks can provide references for further research on conditional sound generation.

Haohe Liu, Qiuqiang Kong, Xubo Liu et al.

This study defines a new evaluation metric for audio tagging tasks to overcome the limitation of the conventional mean average precision (mAP) metric, which treats different kinds of sound as independent classes without considering their relations. Also, due to the ambiguities in sound labeling, the labels in the training and evaluation set are not guaranteed to be accurate and exhaustive, which poses challenges for robust evaluation with mAP. The proposed metric, ontology-aware mean average precision (OmAP) addresses the weaknesses of mAP by utilizing the AudioSet ontology information during the evaluation. Specifically, we reweight the false positive events in the model prediction based on the ontology graph distance to the target classes. The OmAP measure also provides more insights into model performance by evaluations with different coarse-grained levels in the ontology graph. We conduct human evaluations and demonstrate that OmAP is more consistent with human perception than mAP. To further verify the importance of utilizing the ontology information, we also propose a novel loss function (OBCE) that reweights binary cross entropy (BCE) loss based on the ontology distance. Our experiment shows that OBCE can improve both mAP and OmAP metrics on the AudioSet tagging task.

Yi Yuan, Haohe Liu, Xubo Liu et al.

Foley sound generation aims to synthesise the background sound for multimedia content. Previous models usually employ a large development set with labels as input (e.g., single numbers or one-hot vector). In this work, we propose a diffusion model based system for Foley sound generation with text conditions. To alleviate the data scarcity issue, our model is initially pre-trained with large-scale datasets and fine-tuned to this task via transfer learning using the contrastive language-audio pertaining (CLAP) technique. We have observed that the feature embedding extracted by the text encoder can significantly affect the performance of the generation model. Hence, we introduce a trainable layer after the encoder to improve the text embedding produced by the encoder. In addition, we further refine the generated waveform by generating multiple candidate audio clips simultaneously and selecting the best one, which is determined in terms of the similarity score between the embedding of the candidate clips and the embedding of the target text label. Using the proposed method, our system ranks ${1}^{st}$ among the systems submitted to DCASE Challenge 2023 Task 7. The results of the ablation studies illustrate that the proposed techniques significantly improve sound generation performance. The codes for implementing the proposed system are available online.

Haohe Liu, Xuenan Xu, Yi Yuan et al.

Large language models (LLMs) have significantly advanced audio processing through audio codecs that convert audio into discrete tokens, enabling the application of language modelling techniques to audio data. However, traditional codecs often operate at high bitrates or within narrow domains such as speech and lack the semantic clues required for efficient language modelling. Addressing these challenges, we introduce SemantiCodec, a novel codec designed to compress audio into fewer than a hundred tokens per second across diverse audio types, including speech, general sound, and music, without compromising quality. SemantiCodec features a dual-encoder architecture: a semantic encoder using a self-supervised pre-trained Audio Masked Autoencoder (AudioMAE), discretized using k-means clustering on extensive audio data, and an acoustic encoder to capture the remaining details. The semantic and acoustic encoder outputs are used to reconstruct audio via a diffusion-model-based decoder. SemantiCodec is presented in three variants with token rates of 25, 50, and 100 per second, supporting a range of ultra-low bit rates between 0.31 kbps and 1.40 kbps. Experimental results demonstrate that SemantiCodec significantly outperforms the state-of-the-art Descript codec on reconstruction quality. Our results also suggest that SemantiCodec contains significantly richer semantic information than all evaluated state-of-the-art audio codecs, even at significantly lower bitrates. Our code and demos are available at https://haoheliu.github.io/SemantiCodec/.

Arshdeep Singh, Haohe Liu, Mark D. Plumbley

Sounds carry an abundance of information about activities and events in our everyday environment, such as traffic noise, road works, music, or people talking. Recent machine learning methods, such as convolutional neural networks (CNNs), have been shown to be able to automatically recognize sound activities, a task known as audio tagging. One such method, pre-trained audio neural networks (PANNs), provides a neural network which has been pre-trained on over 500 sound classes from the publicly available AudioSet dataset, and can be used as a baseline or starting point for other tasks. However, the existing PANNs model has a high computational complexity and large storage requirement. This could limit the potential for deploying PANNs on resource-constrained devices, such as on-the-edge sound sensors, and could lead to high energy consumption if many such devices were deployed. In this paper, we reduce the computational complexity and memory requirement of the PANNs model by taking a pruning approach to eliminate redundant parameters from the PANNs model. The resulting Efficient PANNs (E-PANNs) model, which requires 36\% less computations and 70\% less memory, also slightly improves the sound recognition (audio tagging) performance. The code for the E-PANNs model has been released under an open source license.

Haohe Liu, Qiuqiang Kong, Jiafeng Liu

Music source separation (MSS) shows active progress with deep learning models in recent years. Many MSS models perform separations on spectrograms by estimating bounded ratio masks and reusing the phases of the mixture. When using convolutional neural networks (CNN), weights are usually shared within a spectrogram during convolution regardless of the different patterns between frequency bands. In this study, we propose a new MSS model, channel-wise subband phase-aware ResUNet (CWS-PResUNet), to decompose signals into subbands and estimate an unbound complex ideal ratio mask (cIRM) for each source. CWS-PResUNet utilizes a channel-wise subband (CWS) feature to limit unnecessary global weights sharing on the spectrogram and reduce computational resource consumptions. The saved computational cost and memory can in turn allow for a larger architecture. On the MUSDB18HQ test set, we propose a 276-layer CWS-PResUNet and achieve state-of-the-art (SoTA) performance on vocals with an 8.92 signal-to-distortion ratio (SDR) score. By combining CWS-PResUNet and Demucs, our ByteMSS system ranks the 2nd on vocals score and 5th on average score in the 2021 ISMIR Music Demixing (MDX) Challenge limited training data track (leaderboard A). Our code and pre-trained models are publicly available at: https://github.com/haoheliu/2021-ISMIR-MSS-Challenge-CWS-PResUNet

Haohe Liu, Qiuqiang Kong, Qiao Tian et al.

Speech restoration aims to remove distortions in speech signals. Prior methods mainly focus on single-task speech restoration (SSR), such as speech denoising or speech declipping. However, SSR systems only focus on one task and do not address the general speech restoration problem. In addition, previous SSR systems show limited performance in some speech restoration tasks such as speech super-resolution. To overcome those limitations, we propose a general speech restoration (GSR) task that attempts to remove multiple distortions simultaneously. Furthermore, we propose VoiceFixer, a generative framework to address the GSR task. VoiceFixer consists of an analysis stage and a synthesis stage to mimic the speech analysis and comprehension of the human auditory system. We employ a ResUNet to model the analysis stage and a neural vocoder to model the synthesis stage. We evaluate VoiceFixer with additive noise, room reverberation, low-resolution, and clipping distortions. Our baseline GSR model achieves a 0.499 higher mean opinion score (MOS) than the speech enhancement SSR model. VoiceFixer further surpasses the GSR baseline model on the MOS score by 0.256. Moreover, we observe that VoiceFixer generalizes well to severely degraded real speech recordings, indicating its potential in restoring old movies and historical speeches. The source code is available at https://github.com/haoheliu/voicefixer_main.

Qiuqiang Kong, Yin Cao, Haohe Liu et al.

Deep neural network based methods have been successfully applied to music source separation. They typically learn a mapping from a mixture spectrogram to a set of source spectrograms, all with magnitudes only. This approach has several limitations: 1) its incorrect phase reconstruction degrades the performance, 2) it limits the magnitude of masks between 0 and 1 while we observe that 22% of time-frequency bins have ideal ratio mask values of over~1 in a popular dataset, MUSDB18, 3) its potential on very deep architectures is under-explored. Our proposed system is designed to overcome these. First, we propose to estimate phases by estimating complex ideal ratio masks (cIRMs) where we decouple the estimation of cIRMs into magnitude and phase estimations. Second, we extend the separation method to effectively allow the magnitude of the mask to be larger than 1. Finally, we propose a residual UNet architecture with up to 143 layers. Our proposed system achieves a state-of-the-art MSS result on the MUSDB18 dataset, especially, a SDR of 8.98~dB on vocals, outperforming the previous best performance of 7.24~dB. The source code is available at: https://github.com/bytedance/music_source_separation

Zhen Ye, Xinfa Zhu, Chi-Min Chan et al.

Recent advances in text-based large language models (LLMs), particularly in the GPT series and the o1 model, have demonstrated the effectiveness of scaling both training-time and inference-time compute. However, current state-of-the-art TTS systems leveraging LLMs are often multi-stage, requiring separate models (e.g., diffusion models after LLM), complicating the decision of whether to scale a particular model during training or testing. This work makes the following contributions: First, we explore the scaling of train-time and inference-time compute for speech synthesis. Second, we propose a simple framework Llasa for speech synthesis that employs a single-layer vector quantizer (VQ) codec and a single Transformer architecture to fully align with standard LLMs such as Llama. Our experiments reveal that scaling train-time compute for Llasa consistently improves the naturalness of synthesized speech and enables the generation of more complex and accurate prosody patterns. Furthermore, from the perspective of scaling inference-time compute, we employ speech understanding models as verifiers during the search, finding that scaling inference-time compute shifts the sampling modes toward the preferences of specific verifiers, thereby improving emotional expressiveness, timbre consistency, and content accuracy. In addition, we released the checkpoint and training code for our TTS model (1B, 3B, 8B) and codec model publicly available.

Ruibin Yuan, Hanfeng Lin, Shuyue Guo et al.

We tackle the task of long-form music generation--particularly the challenging \textbf{lyrics-to-song} problem--by introducing YuE, a family of open foundation models based on the LLaMA2 architecture. Specifically, YuE scales to trillions of tokens and generates up to five minutes of music while maintaining lyrical alignment, coherent musical structure, and engaging vocal melodies with appropriate accompaniment. It achieves this through (1) track-decoupled next-token prediction to overcome dense mixture signals, (2) structural progressive conditioning for long-context lyrical alignment, and (3) a multitask, multiphase pre-training recipe to converge and generalize. In addition, we redesign the in-context learning technique for music generation, enabling versatile style transfer (e.g., converting Japanese city pop into an English rap while preserving the original accompaniment) and bidirectional generation. Through extensive evaluation, we demonstrate that YuE matches or even surpasses some of the proprietary systems in musicality and vocal agility. In addition, fine-tuning YuE enables additional controls and enhanced support for tail languages. Furthermore, beyond generation, we show that YuE's learned representations can perform well on music understanding tasks, where the results of YuE match or exceed state-of-the-art methods on the MARBLE benchmark. Keywords: lyrics2song, song generation, long-form, foundation model, music generation

Zhen Ye, Zeqian Ju, Haohe Liu et al.

Recent progress in large-scale zero-shot speech synthesis has been significantly advanced by language models and diffusion models. However, the generation process of both methods is slow and computationally intensive. Efficient speech synthesis using a lower computing budget to achieve quality on par with previous work remains a significant challenge. In this paper, we present FlashSpeech, a large-scale zero-shot speech synthesis system with approximately 5\% of the inference time compared with previous work. FlashSpeech is built on the latent consistency model and applies a novel adversarial consistency training approach that can train from scratch without the need for a pre-trained diffusion model as the teacher. Furthermore, a new prosody generator module enhances the diversity of prosody, making the rhythm of the speech sound more natural. The generation processes of FlashSpeech can be achieved efficiently with one or two sampling steps while maintaining high audio quality and high similarity to the audio prompt for zero-shot speech generation. Our experimental results demonstrate the superior performance of FlashSpeech. Notably, FlashSpeech can be about 20 times faster than other zero-shot speech synthesis systems while maintaining comparable performance in terms of voice quality and similarity. Furthermore, FlashSpeech demonstrates its versatility by efficiently performing tasks like voice conversion, speech editing, and diverse speech sampling. Audio samples can be found in https://flashspeech.github.io/.

Yi Yuan, Zhuo Chen, Xubo Liu et al.

Contrastive language-audio pretraining~(CLAP) has been developed to align the representations of audio and language, achieving remarkable performance in retrieval and classification tasks. However, current CLAP struggles to capture temporal information within audio and text features, presenting substantial limitations for tasks such as audio retrieval and generation. To address this gap, we introduce T-CLAP, a temporal-enhanced CLAP model. We use Large Language Models~(LLMs) and mixed-up strategies to generate temporal-contrastive captions for audio clips from extensive audio-text datasets. Subsequently, a new temporal-focused contrastive loss is designed to fine-tune the CLAP model by incorporating these synthetic data. We conduct comprehensive experiments and analysis in multiple downstream tasks. T-CLAP shows improved capability in capturing the temporal relationship of sound events and outperforms state-of-the-art models by a significant margin.

Jisheng Bai, Mou Wang, Haohe Liu et al.

Acoustic scene classification (ASC) is a crucial research problem in computational auditory scene analysis, and it aims to recognize the unique acoustic characteristics of an environment. One of the challenges of the ASC task is the domain shift between training and testing data. Since 2018, ASC challenges have focused on the generalization of ASC models across different recording devices. Although this task, in recent years, has achieved substantial progress in device generalization, the challenge of domain shift between different geographical regions, involving discrepancies such as time, space, culture, and language, remains insufficiently explored at present. In addition, considering the abundance of unlabeled acoustic scene data in the real world, it is important to study the possible ways to utilize these unlabelled data. Therefore, we introduce the task Semi-supervised Acoustic Scene Classification under Domain Shift in the ICME 2024 Grand Challenge. We encourage participants to innovate with semi-supervised learning techniques, aiming to develop more robust ASC models under domain shift.

Han Yin, Yang Xiao, Rohan Kumar Das et al.

Audio generation systems now create very realistic soundscapes that can enhance media production, but also pose potential risks. Several studies have examined deepfakes in speech or singing voice. However, environmental sounds have different characteristics, which may make methods for detecting speech and singing deepfakes less effective for real-world sounds. In addition, existing datasets for environmental sound deepfake detection are limited in scale and audio types. To address this gap, we introduce EnvSDD, the first large-scale curated dataset designed for this task, consisting of 45.25 hours of real and 316.74 hours of fake audio. The test set includes diverse conditions to evaluate the generalizability, such as unseen generation models and unseen datasets. We also propose an audio deepfake detection system, based on a pre-trained audio foundation model. Results on EnvSDD show that our proposed system outperforms the state-of-the-art systems from speech and singing domains.

Liumeng Xue, Ziya Zhou, Jiahao Pan et al.

Recent advancements in audio tokenization have significantly enhanced the integration of audio capabilities into large language models (LLMs). However, audio understanding and generation are often treated as distinct tasks, hindering the development of truly unified audio-language models. While instruction tuning has demonstrated remarkable success in improving generalization and zero-shot learning across text and vision, its application to audio remains largely unexplored. A major obstacle is the lack of comprehensive datasets that unify audio understanding and generation. To address this, we introduce Audio-FLAN, a large-scale instruction-tuning dataset covering 80 diverse tasks across speech, music, and sound domains, with over 100 million instances. Audio-FLAN lays the foundation for unified audio-language models that can seamlessly handle both understanding (e.g., transcription, comprehension) and generation (e.g., speech, music, sound) tasks across a wide range of audio domains in a zero-shot manner. The Audio-FLAN dataset is available on HuggingFace and GitHub and will be continuously updated.

Yizhu Jin, Zhen Ye, Zeyue Tian et al.

Diffusion models have demonstrated remarkable success in generative tasks, including audio super-resolution (SR). In many applications like movie post-production and album mastering, substantial computational budgets are available for achieving superior audio quality. However, while existing diffusion approaches typically increase sampling steps to improve quality, the performance remains fundamentally limited by the stochastic nature of the sampling process, leading to high-variance and quality-limited outputs. Here, rather than simply increasing the number of sampling steps, we propose a different paradigm through inference-time scaling for SR, which explores multiple solution trajectories during the sampling process. Different task-specific verifiers are developed, and two search algorithms, including the random search and zero-order search for SR, are introduced. By actively guiding the exploration of the high-dimensional solution space through verifier-algorithm combinations, we enable more robust and higher-quality outputs. Through extensive validation across diverse audio domains (speech, music, sound effects) and frequency ranges, we demonstrate consistent performance gains, achieving improvements of up to 9.70% in aesthetics, 5.88% in speaker similarity, 15.20% in word error rate, and 46.98% in spectral distance for speech SR from 4kHz to 24kHz, showcasing the effectiveness of our approach. Audio samples are available at: https://racerk.github.io/tt-scale-audiosr/.

Junqi Zhao, Jinzheng Zhao, Haohe Liu et al.

Diffusion models have significantly improved the quality and diversity of audio generation but are hindered by slow inference speed. Rectified flow enhances inference speed by learning straight-line ordinary differential equation (ODE) paths. However, this approach requires training a flow-matching model from scratch and tends to perform suboptimally, or even poorly, at low step counts. To address the limitations of rectified flow while leveraging the advantages of advanced pre-trained diffusion models, this study integrates pre-trained models with the rectified diffusion method to improve the efficiency of text-to-audio (TTA) generation. Specifically, we propose AudioTurbo, which learns first-order ODE paths from deterministic noise sample pairs generated by a pre-trained TTA model. Experiments on the AudioCaps dataset demonstrate that our model, with only 10 sampling steps, outperforms prior models and reduces inference to 3 steps compared to a flow-matching-based acceleration model.

Meng Cui, Xianghu Yue, Xinyuan Qian et al.

Fish Feeding Intensity Assessment (FFIA) is crucial in industrial aquaculture management. Recent multi-modal approaches have shown promise in improving FFIA robustness and efficiency. However, these methods face significant challenges when adapting to new fish species or environments due to catastrophic forgetting and the lack of suitable datasets. To address these limitations, we first introduce AV-CIL-FFIA, a new dataset comprising 81,932 labelled audio-visual clips capturing feeding intensities across six different fish species in real aquaculture environments. Then, we pioneer audio-visual class incremental learning (CIL) for FFIA and demonstrate through benchmarking on AV-CIL-FFIA that it significantly outperforms single-modality methods. Existing CIL methods rely heavily on historical data. Exemplar-based approaches store raw samples, creating storage challenges, while exemplar-free methods avoid data storage but struggle to distinguish subtle feeding intensity variations across different fish species. To overcome these limitations, we introduce HAIL-FFIA, a novel audio-visual class-incremental learning framework that bridges this gap with a prototype-based approach that achieves exemplar-free efficiency while preserving essential knowledge through compact feature representations. Specifically, HAIL-FFIA employs hierarchical representation learning with a dual-path knowledge preservation mechanism that separates general intensity knowledge from fish-specific characteristics. Additionally, it features a dynamic modality balancing system that adaptively adjusts the importance of audio versus visual information based on feeding behaviour stages. Experimental results show that HAIL-FFIA is superior to SOTA methods on AV-CIL-FFIA, achieving higher accuracy with lower storage needs while effectively mitigating catastrophic forgetting in incremental fish species learning.

Yi Yuan, Xubo Liu, Haohe Liu et al.

With the development of large-scale diffusion-based and language-modeling-based generative models, impressive progress has been achieved in text-to-audio generation. Despite producing high-quality outputs, existing text-to-audio models mainly aim to generate semantically aligned sound and fall short on precisely controlling fine-grained acoustic characteristics of specific sounds. As a result, users that need specific sound content may find it challenging to generate the desired audio clips. In this paper, we present DreamAudio for customized text-to-audio generation (CTTA). Specifically, we introduce a new framework that is designed to enable the model to identify auditory information from user-provided reference concepts for audio generation. Given a few reference audio samples containing personalized audio events, our system can generate new audio samples that include these specific events. In addition, two types of datasets are developed for training and testing the customized systems. The experiments show that the proposed model, DreamAudio, generates audio samples that are highly consistent with the customized audio features and aligned well with the input text prompts. Furthermore, DreamAudio offers comparable performance in general text-to-audio tasks. We also provide a human-involved dataset containing audio events from real-world CTTA cases as the benchmark for customized generation tasks.

Yilan Dong, Haohe Liu, Qing Wang et al.

Existing 3D Gaussian Splatting (3DGS) methods for hand rendering rely on rigid skeletal motion with an oversimplified non-rigid motion model, which fails to capture fine geometric and appearance details. Additionally, they perform densification based solely on per-point gradients and process poses independently, ignoring spatial and temporal correlations. These limitations lead to geometric detail loss, temporal instability, and inefficient point distribution. To address these issues, we propose HandSplat, a novel Gaussian Splatting-based framework that enhances both fidelity and stability for hand rendering. To improve fidelity, we extend standard 3DGS attributes with implicit geometry and appearance embeddings for finer non-rigid motion modeling while preserving the static hand characteristic modeled by original 3DGS attributes. Additionally, we introduce a local gradient-aware densification strategy that dynamically refines Gaussian density in high-variation regions. To improve stability, we incorporate pose-conditioned attribute regularization to encourage attribute consistency across similar poses, mitigating temporal artifacts. Extensive experiments on InterHand2.6M demonstrate that HandSplat surpasses existing methods in fidelity and stability while achieving real-time performance. We will release the code and pre-trained models upon acceptance.

Xiaofeng Shu, Yehang Zhu, Yanjie Chen et al.

Acoustic echo and background noise can seriously degrade the intelligibility of speech. In practice, echo and noise suppression are usually treated as two separated tasks and can be removed with various digital signal processing (DSP) and deep learning techniques. In this paper, we propose a new cascaded model, magnitude and complex temporal convolutional neural network (MC-TCN), to jointly perform acoustic echo cancellation and noise suppression with the help of adaptive filters. The MC-TCN cascades two separation cores, which are used to extract robust magnitude spectra feature and to enhance magnitude and phase simultaneously. Experimental results reveal that the proposed method can achieve superior performance by removing both echo and noise in real-time. In terms of DECMOS, the subjective test shows our method achieves a mean score of 4.41 and outperforms the INTERSPEECH2021 AEC-Challenge baseline by 0.54.

Qiuqiang Kong, Haohe Liu, Xingjian Du et al.

Speech enhancement is a task to improve the intelligibility and perceptual quality of degraded speech signal. Recently, neural networks based methods have been applied to speech enhancement. However, many neural network based methods require noisy and clean speech pairs for training. We propose a speech enhancement framework that can be trained with large-scale weakly labelled AudioSet dataset. Weakly labelled data only contain audio tags of audio clips, but not the onset or offset times of speech. We first apply pretrained audio neural networks (PANNs) to detect anchor segments that contain speech or sound events in audio clips. Then, we randomly mix two detected anchor segments containing speech and sound events as a mixture, and build a conditional source separation network using PANNs predictions as soft conditions for speech enhancement. In inference, we input a noisy speech signal with the one-hot encoding of "Speech" as a condition to the trained system to predict enhanced speech. Our system achieves a PESQ of 2.28 and an SSNR of 8.75 dB on the VoiceBank-DEMAND dataset, outperforming the previous SEGAN system of 2.16 and 7.73 dB respectively.

Haohe Liu, Lei Xie, Jian Wu et al.

This paper presents a new input format, channel-wise subband input (CWS), for convolutional neural networks (CNN) based music source separation (MSS) models in the frequency domain. We aim to address the major issues in CNN-based high-resolution MSS model: high computational cost and weight sharing between distinctly different bands. Specifically, in this paper, we decompose the input mixture spectra into several bands and concatenate them channel-wise as the model input. The proposed approach enables effective weight sharing in each subband and introduces more flexibility between channels. For comparison purposes, we perform voice and accompaniment separation (VAS) on models with different scales, architectures, and CWS settings. Experiments show that the CWS input is beneficial in many aspects. We evaluate our method on musdb18hq test set, focusing on SDR, SIR and SAR metrics. Among all our experiments, CWS enables models to obtain 6.9% performance gain on the average metrics. With even a smaller number of parameters, less training data, and shorter training time, our MDenseNet with 8-bands CWS input still surpasses the original MMDenseNet with a large margin. Moreover, CWS also reduces computational cost and training time to a large extent.