In recent years, large language models have achieved significant success in generative tasks related to speech, audio, music, and other signal domains. A crucial element of these models is the discrete acoustic codecs, which serve as an intermediate representation replacing the mel-spectrogram. However, there exist several gaps between discrete codecs and downstream speech language models. Specifically, 1) Due to the reconstruction paradigm of the Codec model and the structure of residual vector quantization, the initial channel of the codebooks contains excessive information, making it challenging to directly generate acoustic tokens from weakly supervised signals such as text in downstream tasks. 2) numerous codebooks increases the burden on downstream speech language models. Consequently, leveraging the characteristics of speech language models, we propose Language-Codec. In the Language-Codec, we introduce a Masked Channel Residual Vector Quantization (MCRVQ) mechanism along with improved fourier transform structures and attention blocks, refined discriminator design to address the aforementioned gaps. We compare our method with competing audio compression algorithms and observe significant outperformance across extensive evaluations. Furthermore, we also validate the efficiency of the Language-Codec on downstream speech language models. The source code and pre-trained models can be accessed at https://github.com/jishengpeng/languagecodec .
Neurons in the brain communicate information via punctual events called spikes. The timing of spikes is thought to carry rich information, but it is not clear how to leverage this in digital systems. We demonstrate that event-based encoding is efficient for audio compression. To build this event-based representation we use a deep binary auto-encoder, and under high sparsity pressure, the model enters a regime where the binary event matrix is stored more efficiently with sparse matrix storage algorithms. We test this on the large MAESTRO dataset of piano recordings against vector quantized auto-encoders. Not only does our "Spiking Music compression" algorithm achieve a competitive compression/reconstruction trade-off, but selectivity and synchrony between encoded events and piano key strikes emerge without supervision in the sparse regime.
Many common types of data can be represented as functions that map coordinates to signal values, such as pixel locations to RGB values in the case of an image. Based on this view, data can be compressed by overfitting a compact neural network to its functional representation and then encoding the network weights. However, most current solutions for this are inefficient, as quantization to low-bit precision substantially degrades the reconstruction quality. To address this issue, we propose overfitting variational Bayesian neural networks to the data and compressing an approximate posterior weight sample using relative entropy coding instead of quantizing and entropy coding it. This strategy enables direct optimization of the rate-distortion performance by minimizing the $\beta$-ELBO, and target different rate-distortion trade-offs for a given network architecture by adjusting $\beta$. Moreover, we introduce an iterative algorithm for learning prior weight distributions and employ a progressive refinement process for the variational posterior that significantly enhances performance. Experiments show that our method achieves strong performance on image and audio compression while retaining simplicity.
Language models have been successfully used to model natural signals, such as images, speech, and music. A key component of these models is a high quality neural compression model that can compress high-dimensional natural signals into lower dimensional discrete tokens. To that end, we introduce a high-fidelity universal neural audio compression algorithm that achieves ~90x compression of 44.1 KHz audio into tokens at just 8kbps bandwidth. We achieve this by combining advances in high-fidelity audio generation with better vector quantization techniques from the image domain, along with improved adversarial and reconstruction losses. We compress all domains (speech, environment, music, etc.) with a single universal model, making it widely applicable to generative modeling of all audio. We compare with competing audio compression algorithms, and find our method outperforms them significantly. We provide thorough ablations for every design choice, as well as open-source code and trained model weights. We hope our work can lay the foundation for the next generation of high-fidelity audio modeling.
Mapping music to dance is a challenging problem that requires spatial and temporal coherence along with a continual synchronization with the music's progression. Taking inspiration from large language models, we introduce a 2-step approach for generating dance using a Vector Quantized-Variational Autoencoder (VQ-VAE) to distill motion into primitives and train a Transformer decoder to learn the correct sequencing of these primitives. We also evaluate the importance of music representations by comparing naive music feature extraction using Librosa to deep audio representations generated by state-of-the-art audio compression algorithms. Additionally, we train variations of the motion generator using relative and absolute positional encodings to determine the effect on generated motion quality when generating arbitrarily long sequence lengths. Our proposed approach achieve state-of-the-art results in music-to-motion generation benchmarks and enables the real-time generation of considerably longer motion sequences, the ability to chain multiple motion sequences seamlessly, and easy customization of motion sequences to meet style requirements.
Recent studies show edge computing-based road anomaly detection systems which may also conduct data collection simultaneously. However, the edge computers will have small data storage but we need to store the collected audio samples for a long time in order to update existing models or develop a novel method. Therefore, we should consider an approach for efficient storage management methods while preserving high-fidelity audio. A hardware-perspective approach, such as using a low-resolution microphone, is an intuitive way to reduce file size but is not recommended because it fundamentally cuts off high-frequency components. On the other hand, a computational file compression approach that encodes collected high-resolution audio into a compact code should be recommended because it also provides a corresponding decoding method. Motivated by this, we propose a way of simple yet effective pre-trained autoencoder-based data compression method. The pre-trained autoencoder is trained for the purpose of audio super-resolution so it can be utilized to encode or decode any arbitrary sampling rate. Moreover, it will reduce the communication cost for data transmission from the edge to the central server. Via the comparative experiments, we confirm that the zero-shot audio compression and decompression highly preserve anomaly detection performance while enhancing storage and transmission efficiency.
Implicit Neural Representations (INRs) have emerged as a promising method for representing diverse data modalities, including 3D shapes, images, and audio. While recent research has demonstrated successful applications of INRs in image and 3D shape compression, their potential for audio compression remains largely unexplored. Motivated by this, we present a preliminary investigation into the use of INRs for audio compression. Our study introduces Siamese SIREN, a novel approach based on the popular SIREN architecture. Our experimental results indicate that Siamese SIREN achieves superior audio reconstruction fidelity while utilizing fewer network parameters compared to previous INR architectures.
Spatial audio quality is a highly multifaceted concept, with many interactions between environmental, geometrical, anatomical, psychological, and contextual considerations. Methods for characterization or evaluation of the geometrical components of spatial audio quality, however, remain scarce, despite being perhaps the least subjective aspect of spatial audio quality to quantify. By considering interchannel time and level differences relative to a reference signal, it is possible to construct a signal model to isolate some of the spatial distortion. By using a combination of least-square optimization and heuristics, we propose a signal decomposition method to isolate the spatial error from a processed signal, in terms of interchannel gain leakages and changes in relative delays. This allows the computation of simple energy-ratio metrics, providing objective measures of spatial and non-spatial signal qualities, with minimal assumptions and no dataset dependency. Experiments demonstrate the robustness of the method against common spatial signal degradation introduced by, e.g., audio compression and music source separation. Implementation is available at https://github.com/karnwatcharasupat/spauq.
Despite the huge successes made in neutral TTS, content-leakage remains a challenge. In this paper, we propose a new input representation and simple architecture to achieve improved prosody modeling. Inspired by the recent success in the use of discrete code in TTS, we introduce discrete code to the input of the reference encoder. Specifically, we leverage the vector quantizer from the audio compression model to exploit the diverse acoustic information it has already been trained on. In addition, we apply the modified MLP-Mixer to the reference encoder, making the architecture lighter. As a result, we train the prosody transfer TTS in an end-to-end manner. We prove the effectiveness of our method through both subjective and objective evaluations. We demonstrate that the reference encoder learns better speaker-independent prosody when discrete code is utilized as input in the experiments. In addition, we obtain comparable results even when fewer parameters are inputted.
In audio processing applications, the generation of expressive sounds based on high-level representations demonstrates a high demand. These representations can be used to manipulate the timbre and influence the synthesis of creative instrumental notes. Modern algorithms, such as neural networks, have inspired the development of expressive synthesizers based on musical instrument timbre compression. Unsupervised deep learning methods can achieve audio compression by training the network to learn a mapping from waveforms or spectrograms to low-dimensional representations. This study investigates the use of stacked convolutional autoencoders for the compression of time-frequency audio representations for a variety of instruments for a single pitch. Further exploration of hyper-parameters and regularization techniques is demonstrated to enhance the performance of the initial design. In an unsupervised manner, the network is able to reconstruct a monophonic and harmonic sound based on latent representations. In addition, we introduce an evaluation metric to measure the similarity between the original and reconstructed samples. Evaluating a deep generative model for the synthesis of sound is a challenging task. Our approach is based on the accuracy of the generated frequencies as it presents a significant metric for the perception of harmonic sounds. This work is expected to accelerate future experiments on audio compression using neural autoencoders.
The rapid spread of media content synthesis technology and the potentially damaging impact of audio and video deepfakes on people's lives have raised the need to implement systems able to detect these forgeries automatically. In this work we present a novel approach for synthetic speech detection, exploiting the combination of two high-level semantic properties of the human voice. On one side, we focus on speaker identity cues and represent them as speaker embeddings extracted using a state-of-the-art method for the automatic speaker verification task. On the other side, voice prosody, intended as variations in rhythm, pitch or accent in speech, is extracted through a specialized encoder. We show that the combination of these two embeddings fed to a supervised binary classifier allows the detection of deepfake speech generated with both Text-to-Speech and Voice Conversion techniques. Our results show improvements over the considered baselines, good generalization properties over multiple datasets and robustness to audio compression.
We introduce a state-of-the-art real-time, high-fidelity, audio codec leveraging neural networks. It consists in a streaming encoder-decoder architecture with quantized latent space trained in an end-to-end fashion. We simplify and speed-up the training by using a single multiscale spectrogram adversary that efficiently reduces artifacts and produce high-quality samples. We introduce a novel loss balancer mechanism to stabilize training: the weight of a loss now defines the fraction of the overall gradient it should represent, thus decoupling the choice of this hyper-parameter from the typical scale of the loss. Finally, we study how lightweight Transformer models can be used to further compress the obtained representation by up to 40%, while staying faster than real time. We provide a detailed description of the key design choices of the proposed model including: training objective, architectural changes and a study of various perceptual loss functions. We present an extensive subjective evaluation (MUSHRA tests) together with an ablation study for a range of bandwidths and audio domains, including speech, noisy-reverberant speech, and music. Our approach is superior to the baselines methods across all evaluated settings, considering both 24 kHz monophonic and 48 kHz stereophonic audio. Code and models are available at github.com/facebookresearch/encodec.
Melodic similarity measurement is of key importance in music information retrieval. In this paper, we use geometric matching techniques to measure the similarity between two melodies. We represent music as sets of points or sets of horizontal line segments in the Euclidean plane and propose efficient algorithms for optimization problems inspired in two operations on melodies; linear scaling and audio compression. In the scaling problem, an incoming query melody is scaled forward until the similarity measure between the query and a reference melody is minimized. The compression problem asks for a subset of notes of a given melody such that the matching cost between the selected notes and the reference melody is minimized.
Previous DCASE challenges contributed to an increase in the performance of acoustic scene classification systems. State-of-the-art classifiers demand significant processing capabilities and memory which is challenging for resource-constrained mobile or IoT edge devices. Thus, it is more likely to deploy these models on more powerful hardware and classify audio recordings previously uploaded (or streamed) from low-power edge devices. In such scenario, the edge device may apply perceptual audio coding to reduce the transmission data rate. This paper explores the effect of perceptual audio coding on the classification performance using a DCASE 2020 challenge contribution [1]. We found that classification accuracy can degrade by up to 57% compared to classifying original (uncompressed) audio. We further demonstrate how lossy audio compression techniques during model training can improve classification accuracy of compressed audio signals even for audio codecs and codec bitrates not included in the training process.
With active research in audio compression techniques yielding substantial breakthroughs, spectral reconstruction of low-quality audio waves remains a less indulged topic. In this paper, we propose a novel approach for reconstructing higher frequencies from considerably longer sequences of low-quality MP3 audio waves. Our technique involves inpainting audio spectrograms with residually stacked autoencoder blocks by manipulating individual amplitude and phase values in relation to perceptual differences. Our architecture presents several bottlenecks while preserving the spectral structure of the audio wave via skip-connections. We also compare several task metrics and demonstrate our visual guide to loss selection. Moreover, we show how to leverage differential quantization techniques to reduce the initial model size by more than half while simultaneously reducing inference time, which is crucial in real-world applications.
In this paper, we present UR-AIR system submission to the logical access (LA) and the speech deepfake (DF) tracks of the ASVspoof 2021 Challenge. The LA and DF tasks focus on synthetic speech detection (SSD), i.e. detecting text-to-speech and voice conversion as spoofing attacks. Different from previous ASVspoof challenges, the LA task this year presents codec and transmission channel variability, while the new task DF presents general audio compression. Built upon our previous research work on improving the robustness of the SSD systems to channel effects, we propose a channel-robust synthetic speech detection system for the challenge. To mitigate the channel variability issue, we use an acoustic simulator to apply transmission codec, compression codec, and convolutional impulse responses to augmenting the original datasets. For the neural network backbone, we propose to use Emphasized Channel Attention, Propagation and Aggregation Time Delay Neural Networks (ECAPA-TDNN) as our primary model. We also incorporate one-class learning with channel-robust training strategies to further learn a channel-invariant speech representation. Our submission achieved EER 20.33% in the DF task; EER 5.46% and min-tDCF 0.3094 in the LA task.
Recent achievements in end-to-end deep learning have encouraged the exploration of tasks dealing with highly structured data with unified deep network models. Having such models for compressing audio signals has been challenging since it requires discrete representations that are not easy to train with end-to-end backpropagation. In this paper, we present an end-to-end deep learning approach that combines recurrent neural networks (RNNs) within the training strategy of variational autoencoders (VAEs) with a binary representation of the latent space. We apply a reparametrization trick for the Bernoulli distribution for the discrete representations, which allows smooth backpropagation. In addition, our approach allows the separation of the encoder and decoder, which is necessary for compression tasks. To our best knowledge, this is the first end-to-end learning for a single audio compression model with RNNs, and our model achieves a Signal to Distortion Ratio (SDR) of 20.54.
The onset of coronavirus disease 2019 (COVID-19), an infectious disease caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), has sparked unprecedented change. Due to the public health guidelines imposed during the COVID-19 pandemic, there is no longer sufficient street traffic for remaining buskers to generate sufficient revenue, leading a majority of street musicians to pursue remote music production. However, real-time music production is notoriously difficult due to the excessively high latencies that current video call platforms such as Zoom and Google Meet harbor. In this paper, we propose an architecture for a platform with end-to-end, near-lossless audio transmission tailored specifically to online joint music production, called Latent Space. We discuss the usage of a recurrent autoencoder with sequence-aware encoding (RAES) and a 1D convolutional layer for audio compression, which we dub ClefNet, as well as propose a new evaluation metric for naive autoencoders (AEs), MSE-DTW loss, which combines the traditional mean square error (MSE) loss function with dynamic time warping (DTW) to prevent an increase in loss when the target sequence predicted by the AE is strictly a temporal variation of the source sequence. Moreover, we detail the logistics of a live system implementation which uses the Web Audio API to extract raw audio samples in real-time to feed into our client-side model before relaying the traffic using peer-to-peer WebRTC technology. The Latent Space platform can be accessed at https://latent-space.tech, and the code and data can be found under the MIT License at https://github.com/rvignav/ClefNet.
We present a deep convolutional GAN which leverages techniques from MP3/Vorbis audio compression to produce long, high-quality audio samples with long-range coherence. The model uses a Modified Discrete Cosine Transform (MDCT) data representation, which includes all phase information. Phase generation is hence integral part of the model. We leverage the auditory masking and psychoacoustic perception limit of the human ear to widen the true distribution and stabilize the training process. The model architecture is a deep 2D convolutional network, where each subsequent generator model block increases the resolution along the time axis and adds a higher octave along the frequency axis. The deeper layers are connected with all parts of the output and have the context of the full track. This enables generation of samples which exhibit long-range coherence. We use MP3net to create 95s stereo tracks with a 22kHz sample rate after training for 250h on a single Cloud TPUv2. An additional benefit of the CNN-based model architecture is that generation of new songs is almost instantaneous.
In a number of data-driven applications such as detection of arrhythmia, interferometry or audio compression, observations are acquired indistinctly in the time or frequency domains: temporal observations allow us to study the spectral content of signals (e.g., audio), while frequency-domain observations are used to reconstruct temporal/spatial data (e.g., MRI). Classical approaches for spectral analysis rely either on i) a discretisation of the time and frequency domains, where the fast Fourier transform stands out as the \textit{de facto} off-the-shelf resource, or ii) stringent parametric models with closed-form spectra. However, the general literature fails to cater for missing observations and noise-corrupted data. Our aim is to address the lack of a principled treatment of data acquired indistinctly in the temporal and frequency domains in a way that is robust to missing or noisy observations, and that at the same time models uncertainty effectively. To achieve this aim, we first define a joint probabilistic model for the temporal and spectral representations of signals, to then perform a Bayesian model update in the light of observations, thus jointly reconstructing the complete (latent) time and frequency representations. The proposed model is analysed from a classical spectral analysis perspective, and its implementation is illustrated through intuitive examples. Lastly, we show that the proposed model is able to perform joint time and frequency reconstruction of real-world audio, healthcare and astronomy signals, while successfully dealing with missing data and handling uncertainty (noise) naturally against both classical and modern approaches for spectral estimation.