Fast and Flexible Audio Bandwidth Extension via Vocos

Yatharth Sharma

eess.ASMar 7, 2026

Fast and Flexible Audio Bandwidth Extension via Vocos

Authors: Yatharth Sharma

Abstract

We propose a Vocos-based bandwidth extension model that enhances audio at 8-48 kHz by generating missing high-frequency content. Inputs are resampled to 48 kHz and processed by a neural vocoder backbone, enabling a single network to support arbitrary upsampling ratios. A lightweight Linkwitz-Riley-inspired refiner merges the original low band with the generated high frequencies via a smooth crossover. On validation, the model achieves competitive log-spectral distance while running at a real-time factor of 0.0001 on an NVIDIA A100 GPU and 0.0053 on an 8-core CPU, demonstrating practical, high-quality BWE at extreme throughput.

Figures & tables

Figure 1: Overview of the proposed BWE architecture following the Vocos framework. The input audio is upsampled to 48 kHz via Sinc interpolation and transformed into an 80-bin Mel-spectrogram. The backbone consists of 8 ConvNeXt blocks using $7\times 1$ depthwise and $1\times 1$ pointwise convolutions. The head employs a linear layer and ISTFT for waveform reconstruction, followed by a Linkwitz-Riley inspired refiner for significant quality enhancement.

Method	8 $\rightarrow$ 48 kHz (LSD $\downarrow$ )	12 $\rightarrow$ 48 kHz (LSD $\downarrow$ )	16 $\rightarrow$ 48 kHz (LSD $\downarrow$ )
Sinc upsampling	3.52	3.19	2.94
AudioSR	1.61	1.52	1.44
NVSR	1.22	1.20	1.12
AP-BWE	0.87	0.81	0.74
Proposed Model	0.85	0.80	0.74

Table 1: Log-spectral distance (LSD) comparison on VCTK. Our model achieves very competitive spectral reconstruction across all tiers.

Method	8 $\rightarrow$ 48 kHz (ViSQOL $\uparrow$ )	12 $\rightarrow$ 48 kHz (ViSQOL $\uparrow$ )	16 $\rightarrow$ 48 kHz (ViSQOL $\uparrow$ )
Sinc upsampling	2.10	2.23	2.35
AudioSR	3.15	3.21	3.40
NVSR	2.97	3.00	3.10
AP-BWE	3.51	3.54	3.70
Proposed Model	3.51	3.53	3.69

Table 2: ViSQOL scores on VCTK. Our model is similar to AP-BWE’s perceptual quality.

Panel of Figure 2 — Figure 2: Model performance (LSD) across in-domain and out-of-domain (OOD) sample rates. The curve demonstrates a linear improvement in fidelity as input bandwidth increases, regardless of whether the rate was seen during training.

Refiner Variant	LSD $\downarrow$ (8 $\rightarrow$ 48)
No Refiner (Direct Output)	0.897
Naive HP/LP Crossover	0.865
Standard Butterworth Refiner	0.861
Linkwitz-Riley Inspired (Proposed)	0.850

Table 3: Ablation results on VCTK validation. The proposed refiner achieves the lowest spectral distance by ensuring a flat magnitude response at the crossover.

Method	Params	RTF $\downarrow$	Speed $\uparrow$
AudioSR	258M	50.0000	0.02 $\times$
NVSR	$\sim$ 99M	0.2218	4.5 $\times$
AP-BWE	3*30M	0.0495	20.2 $\times$
Proposed	15M	0.0053	190.5 $\times$

Table 4: Inference efficiency comparison on 8-core CPU. RTF is measured as

(latency/duration)

\text{Speed}=1/\text{RTF}

Method	BS	Latency (s)	RTF $\downarrow$	Speed $\uparrow$
AudioSR	1	8.4700	2.1175	0.47 $\times$
NVSR	1	0.0410	0.0103	97.6 $\times$
AP-BWE	1	0.0135	0.0034	296.3 $\times$
AP-BWE	32	0.2879	0.0023	444.6 $\times$
Proposed	1	0.0025	0.0006	1600.0 $\times$
Proposed	32	0.0102	0.0001	12549.0 $\times$

Table 5: GPU Efficiency and throughput comparison on NVIDIA A100. Latency is reported for the entire batch. RTF is rounded and calculated as

latency/(duration\times BS)

Neural Vocoders

See all 9 papers

Multi-Rate Bandwidth Extension by Token Completion in Neural Audio Codecs
Sep 29, 2026Benoît Ginies, Olivier Fercoq, Gaël RichardAudio Representation Learning Neural Audio Codecs
LTCI, Télécom Paris, Institut Polytechnique de Paris, Palaiseau, France
AnyBand: Unified Multi-Bandwidth Speech Extension via Frequency-Aware In-Context Spectral Infilling
Aug 1, 2026Junchuan Zhao, Minh Duc Vu, Bowen Zhang +1Audio Diffusion Models Speech Enhancement
School of Computing, National University of Singapore · Department of Statistics & Data Science, National University of Singapore · College of Computing and Data Science, Nanyang Technological University
Harmonic-Percussive Disentangled Neural Audio Codec for Bandwidth Extension
Nov 26, 2025Benoît Giniès, Xiaoyu Bie, Olivier Fercoq +1Audio Representation Learning Neural Audio Codecs
LTCI, Télécom Paris, Institut Polytechnique de Paris Palaiseau, France

Sep 29, 2026J/K move · Enter open · S save

Multi-Rate Bandwidth Extension by Token Completion in Neural Audio Codecs

Benoît Ginies, Olivier Fercoq, Gaël Richard

LTCI, Télécom Paris, Institut Polytechnique de Paris, Palaiseau, France

Audio Representation Learning Neural Audio Codecs Audio Tokenization

Open paper AskPDF arXiv

Bandwidth extension, the task of reconstructing the high-frequency components of an audio signal from its low-passed counterpart, is a long-standing problem in audio processing. In this work, we extend recent advances in neural architectures by framing bandwidth extension as an audio token prediction problem. Specifically, we train a transformer-based language model on the discrete representations produced by a disentangled neural audio codec, where the disentanglement is guided by a Harmonic-Percussive decomposition of the input signals, highlighting spectral structures particularly relevant for bandwidth extension. Our approach introduces a novel codec design that explicitly accounts for the downstream token prediction task, enabling a more effective coupling between codec structure and transformer modeling. This joint design yields high-quality reconstructions of the original signal, as measured by both objective metrics and subjective evaluations. These results highlight the importance of aligning codec disentanglement and representation learning with the generative modeling stage, and demonstrate the potential of global, representation-aware design for advancing bandwidth extension.

Abstract

Neural Vocoders

Explore similar work

Multi-Rate Bandwidth Extension by Token Completion in Neural Audio Codecs

AnyBand: Unified Multi-Bandwidth Speech Extension via Frequency-Aware In-Context Spectral Infilling

Harmonic-Percussive Disentangled Neural Audio Codec for Bandwidth Extension