An official website of the United States government
Neural speech codecs have revolutionized speech coding, achieving higher compression while preserving audio fidelity. Beyond compression, they have emerged as tokenization strategies, enabling language modeling on speech and driving paradigm shifts across various speech processing tasks. Despite these advancements, their robustness in noisy environments remains underexplored, raising concerns about their generalization to real-world scenarios. In this work, we systematically evaluate neural speech codecs under various noise conditions, revealing non-trivial differences in their robustness. We further examine their linearity properties, uncovering non-linear distortions which partly explain observed variations in robustness. Lastly, we analyze their frequency response to identify factors affecting audio fidelity. Our findings provide critical insights into codec behavior and future codec design, as well as emphasizing the importance of noise robustness for their real-world integration.
more »
« less
This content will become publicly available on April 6, 2026
Code Drift: Towards Idempotent Neural Audio Codecs
Neural codecs have demonstrated strong performance in high-fidelity compression of audio signals at low bitrates. The token-based representations produced by these codecs have proven particularly useful for generative modeling. While much research has focused on improvements in compression ratio and perceptual transparency, recent works have largely overlooked another desirable codec property -- \textit{idempotence}, the stability of compressed outputs under multiple rounds of encoding. We find that state-of-the-art neural codecs exhibit varied degrees of idempotence, with some degrading audio outputs significantly after as few as three encodings. We investigate possible causes of low idempotence and devise a method for improving idempotence through fine-tuning a codec model. We then examine the effect of idempotence on a simple conditional generative modeling task, and find that increased idempotence can be achieved without negatively impacting downstream modeling performance -- potentially extending the usefulness of neural codecs for practical file compression and iterative generative modeling workflows.
more »
« less
- Award ID(s):
- 2222369
- PAR ID:
- 10638310
- Publisher / Repository:
- IEEE
- Date Published:
- ISBN:
- 979-8-3503-6874-1
- Page Range / eLocation ID:
- 1 to 5
- Subject(s) / Keyword(s):
- generative models tokenization codec
- Format(s):
- Medium: X
- Location:
- Hyderabad, India
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
With the expansion of sensor nodes to newer avenues of technologies, such as the Internet of things (IoT), internet of bodies (IoB), augmented reality (AR), and mixed reality, the demand to support high-speed operations, such as audio and video, with a minimal increase in power consumption is gaining much traction. In this work, we focus on these nodes operating in audio-based AR (AAR) and explore the opportunity of supporting audio at a low power budget. For sensor nodes, communicating one bit of data usually consumes significantly higher power than the power associated with sensing and processing/computing one data bit. Compressing the number of communication bits at the expense of a few computation cycles considerably reduces the overall power consumption of the nodes. Audio codecs such as AAC and LDAC that currently perform compression and decompression of audio streams burn significant power and create a floor to the minimum power possible in these applications. Compressive sensing (CS), a powerful mathematical tool for compression, is often used in physiological signal sensing, such as EEG and ECG, and it can offer a promising low-power alternative to audio codecs. We introduce a new paradigm of using the CS-based approach to realize audio compression that can function as a new independent technique or augment the existing codecs for a higher level of compression. This work, CS-Audio, fabricated in TSMC 65-nm CMOS technology, presents the first CS-based compression, equipped with an ON-chip DWT sparsifier for non-sparse audio signals. The CS design, realized in a pipelined architecture, achieves high data rates and enables a wake-up implementation to bypass computation for insignificant input samples, reducing the power consumption of the hardware. The measurement results demonstrate a 3X-15X reduction in transmitted audio data without a perceivable degradation of audio quality, as indicated by the perceptual evaluation of audio quality mean opinion score (PEAQ MOS) >1.5. The hardware consumes 238 μW power at 0.65 V and 15 Mbps, which is (~20X-40X) lower than audio codecs.more » « less
-
The past decade has witnessed the rising dominance of deep learning and artificial intelligence in a wide range of applications. In particular, the ocean of wireless smartphones and IoT devices continue to fuel the tremendous growth of edge/cloudbased machine learning (ML) systems including image/speech recognition and classification. To overcome the infrastructural barrier of limited network bandwidth in cloud ML, existing solutions have mainly relied on traditional compression codecs such as JPEG that were historically engineered for humanend users instead of ML algorithms. Traditional codecs do not necessarily preserve features important to ML algorithms under limited bandwidth, leading to potentially inferior performance. This work investigates application-driven optimization of programmable commercial codec settings for networked learning tasks such as image classification. Based on the foundation of variational autoencoders (VAEs), we develop an end-to-end networked learning framework by jointly optimizing the codec and classifier without reconstructing images for given data rate (bandwidth). Compared with standard JPEG codec, the proposed VAE joint compression and classification framework achieves classification accuracy improvement by over 10% and 4%, respectively, for CIFAR-10 and ImageNet-1k data sets at data rate of 0.8 bpp. Our proposed VAE-based models show 65%99% reductions in encoder size, 1.5 13.1 improvements in inference speed and 25%99% savings in power compared to baseline models. We further show that a simple decoder can reconstruct images with sufficient quality without compromising classification accuracy.more » « less
-
null (Ed.)Compressed videos constitute 70% of Internet traffic, and video upload growth rates far outpace compute and storage improvement trends. Past work in leveraging perceptual cues like saliency, i.e., regions where viewers focus their perceptual attention, reduces compressed video size while maintaining perceptual quality, but requires significant changes to video codecs and ignores the data management of this perceptual information. In this paper, we propose Vignette, a compression technique and storage manager for perception-based video compression in the cloud. Vignette complements off-the-shelf compression software and hardware codec implementations. Vignette's compression technique uses a neural network to predict saliency information used during transcoding, and its storage manager integrates perceptual information into the video storage system. Our results demonstrate the benefit of embedding information about the human visual system into the architecture of cloud video storage systems.more » « less
-
The emergence of Audio-based Augmented Reality has been calling for increasing data-rates for audio signals, with significant reduction in power to enable extremely energy-constrained sensor nodes. Typically, the communication power dominates sensing and computing power in a node [1]. For highly energy constrained scenarios, compressive sensing (CS) have been demonstrated (Fig. 1), where samples are first compressed at the sensor to contain the same information in a smaller number of samples, before transmitting to a receiver, where the signal is reconstructed. Previous CS works [2]-[5] have focused entirely on “sparse” physiological signals, operating in low speed regime. This work illustrates the first CS design, enabled with a discrete wavelet transform (DWT) sparsifier for catering to non-sparse signals such as high definition audio. Audio recording and playback are quite sensitive to quality, thereby requiring audio codecs, such as. aac, for efficient compression and decompression of audio streams, which usually consume power in the order of mW [6]. Audio inferencing operated in intelligent assistants are more tolerant to input quality, functioning effectively when the Perceptual Evaluation of Audio Quality Mean Opinion Score (PAEQ MOS) [7], an ITU-R standard objective metric for characterizing perceived audio quality, exceeds 1.5. CS presents an opportunity to achieve >10X reduction in transmitted audio data with orders of magnitude lower power, as compared to codecs. The design is implemented in 65 nm CMOS and consumes 238 uW power at 0.65 V and 15 Mbps.more » « less
