VoiceBank+DEMAND Speech Enhancement Benchmark

• VoiceBank+DEMAND is a composite benchmark that merges clean VoiceBank/VCTK speech with real-world DEMAND noise to simulate controlled single-channel speech enhancement challenges.
• The dataset employs systematic additive mixing at defined SNRs (0, 5, 10, 15 dB in training and 2.5–17.5 dB in testing) to ensure consistent evaluation protocols.
• Standardized splits, resampling, and objective metrics like PESQ and STOI underlie its role in benchmarking advanced deep-learning models for speech enhancement.

The VoiceBank+DEMAND dataset is a composite speech enhancement benchmark that couples clean, read English speech from the Voice Bank/VCTK corpus with diverse, real-world environmental noise from the DEMAND database, systematically combined at specified signal-to-noise ratio (SNR) levels. This setup is rigorously employed as a principal evaluation and training protocol for modern single-channel speech enhancement (SE) systems, including discriminative DNNs, state-space architectures, GANs, diffusion models, and other advanced methods.

1. Dataset Composition and Structure

VoiceBank+DEMAND consists of two principal elements: clean-speech signals and environmental noise sources. The clean-speech subset is drawn from the VoiceBank (VCTK) corpus and is typically downsampled from 48 kHz to 16 kHz for computational efficiency and PESQ compatibility. Across most studies, the standard partition contains 28 training speakers and 2 test speakers, though some works introduce a further validation split from the training population, frequently holding out two speakers for this purpose (Ku et al., 2023, Li et al., 23 Dec 2024, Chao et al., 2022, Park et al., 2022, Wen et al., 20 Aug 2025, Kühne et al., 10 Jan 2025, Yen et al., 2022).

Split	#Speakers	#Clean Utterances	#Noise Types	SNRs (dB)
Train	28	~10,802 – 11,572	8 DEMAND + 2 artificial	0, 5, 10, 15
Validation	2	~770 – 800	as train	as train
Test	2	824	5 unseen DEMAND	2.5, 7.5, 12.5, 17.5

Clean utterances from the training and validation splits collectively amount to 11,572 examples in most configurations, with an average duration of 3–5 seconds per utterance, providing close to 12 hours of speech. The test set comprises 824 (occasionally 872) utterances sourced from two previously unseen speakers. The gender distribution is sometimes controlled for, with explicit balancing (14 male, 14 female) in certain works (Park et al., 2022).

Noise is introduced via ten environmental categories: typically eight authentic DEMAND recordings (e.g., street, café, kitchen, pedestrian, car, bus, etc.) plus two artificial sources (e.g., white and pink noise). Test noise always consists of held-out DEMAND environments.

2. Mixing and Augmentation Protocols

Noisy mixtures are generated through additive mixing of clean and noise segments, where the noise is scaled to a prescribed SNR relative to the clean signal's RMS energy. The canonical formula is

$\text{SNR(dB)} = 10 \log_{10}(P_x / P_n)$

$y(t) = x(t) + \alpha n(t)$

with $\alpha$ chosen to achieve the target SNR (Ku et al., 2023, Li et al., 23 Dec 2024, Park et al., 2022).

Training SNRs are uniformly distributed among $\{0, 5, 10, 15\}$ dB. Testing is conducted at offset SNRs $\{2.5, 7.5, 12.5, 17.5\}$ dB, ensuring challenging generalization. Segment selection and starting points for noise are randomized for every mixture creation to avoid overfitting.

Several studies augment this protocol with advanced regularization and data augmentation methods such as amplitude normalization over complex spectrograms, the Remix technique (intra-batch shuffling of noise), and BandMask (randomly removing 20% of spectral bands) (Ku et al., 2023).

3. Preprocessing and Data Representation

All speech and noise materials are resampled to 16 kHz and formatted as 16-bit PCM WAV files. Standardized file and directory structures facilitate systematic evaluation, with metadata linking each noisy utterance to its clean counterpart and the specific noise/SNR configuration employed (Park et al., 2022).

Feature extraction for most time-frequency (TF) domain systems relies on short-time Fourier transforms (STFT). Window lengths of 400 samples, 400-sample FFT sizes, and 100-sample hops with Hann windows are prevalent in leading systems (Kühne et al., 10 Jan 2025). Magnitude compression, such as power-law scaling $|Y_m|^c$ (common $c=0.3$), is extensively adopted to stabilize mask-inference or to regularize deep SE models (Wen et al., 20 Aug 2025, Kühne et al., 10 Jan 2025).

4. Evaluation Metrics and Benchmarks

The dataset’s primary role is as a reproducible and challenging testbed for single-channel SE models. The following objective metrics are universally computed:

• PESQ (Perceptual Evaluation of Speech Quality; ITU-T P.862, range $\sim$[–0.5, 4.5])
• STOI (Short-Time Objective Intelligibility; range [0, 1])
• CSIG (MOS: signal distortion, [1, 5])
• CBAK (MOS: background noise intrusiveness, [1, 5])
• COVL (MOS: overall quality, [1, 5])
• SSNR (Segmental SNR; dB scale)

Typical baselines include:

• No enhancement: PESQ $\sim$1.97
• Classic Wiener filter: 2.22
• Early GANs (SEGAN): 2.16
• Typical discriminative DNNs (Conv-TasNet: 2.53, Demucs: 2.93–3.07)
• State-of-the-art SE (MP-SENet, DPT-FSNet, MetricGAN+, CMGAN, xLSTM-SENet, EffiFusion-GAN): 3.33–3.50 PESQ, with recent models achieving CSIG$>$4.7, COVL$>$4.1, STOI$>$0.96 (Wen et al., 20 Aug 2025, Kühne et al., 10 Jan 2025, Li et al., 23 Dec 2024, Chao et al., 2022, Ku et al., 2023).

5. Variants, Reproducibility, and Research Impact

VoiceBank+DEMAND underpins most state-of-the-art SE development in the last half-decade. Its rigorously controlled protocol (explicit speaker splits, consistent noise/SNR designations, and well-documented data access patterns) has enabled reproducible SE research and fair benchmarking across model classes, including:

• Structured state spaces (S4-SE): substantial model size reduction with competitive PESQ given amplitude normalization, remixing, and BandMask augmentations (Ku et al., 2023).
• Kolmogorov-Arnold Networks (KAN), GR-KAN: up to 4x parameter savings and PESQ improvements of 0.1 using advanced activation parameterizations (Li et al., 23 Dec 2024).
• GAN variants (EffiFusion-GAN): compressed architectures with depthwise convolution, yielding PESQ 3.45 at 1.08M parameters (Wen et al., 20 Aug 2025).
• xLSTM-based systems (xLSTM-SENet): bidirectional TF-xLSTM yielding parity or gains over Mamba- and Conformer-based SE in both accuracy and efficiency (Kühne et al., 10 Jan 2025).
• Attention-based and dual-path networks: structured validation splits, transparent metrics, and explicit file-lists encourage rigorous ablation and cross-paper comparisons (Park et al., 2022).
• Diffusion-based enhancement (cold diffusion): unfolded training with DCCRN backbone closing most of the discriminative–diffusion gap (Yen et al., 2022).

6. Limitations and Usage Considerations

Despite its widespread adoption, VoiceBank+DEMAND is limited by fixed text content (read speech), absence of strong reverberation or far-field effects, and mismatch with conversational or accented spontaneous speech. Noise diversity, while high, is circumscribed to ten environments. Test speakers and test noise environments are always disjoint from training, but neither are especially large. Some works do not specify or document their precise validation protocols, potentially leading to modest inconsistencies in early stopping or hyperparameter selection (Wen et al., 20 Aug 2025).

A plausible implication is that results may not generalize from this benchmark to truly out-of-domain or far-field noise conditions, yet the carefully documented structure makes the dataset ideal for reproducibility and controlled ablation within the prevailing paradigms of SE model research.

7. Summary Table: Canonical VoiceBank+DEMAND Configuration Across Key Works

Paper (arXiv)	Train Spk/Utt	Test Spk/Utt	Noise Types (Train/Test)	SNR Train/Test (dB)
(Ku et al., 2023)	28 / 10,802	2 / 824	10 / 10	0,5,10,15/2.5–17.5
(Li et al., 23 Dec 2024)	28 / 11,572	2 / 824	10 / 5	0,5,10,15/2.5–17.5
(Chao et al., 2022)	28 / 11,572	2 / 824	10 / 10	0,5,10,15/2.5–17.5
(Park et al., 2022)	28 / 11,572	2 / 824	10 / 10	0,5,10,15/2.5–17.5
(Wen et al., 20 Aug 2025)	28 / 11,572	2 / 872	10 / 5	(train SNRs not listed)/2.5–17.5
(Kühne et al., 10 Jan 2025)	28 / 11,572	2 / 824	10 / 10	0,5,10,15/2.5–17.5
(Yen et al., 2022)	28 / 10,802	2 / 824	10 / 10	0,5,10,15/2.5–17.5

This canonicalized benchmark continues to serve as the principal proving ground for contemporary, deep-learning-based single-channel speech enhancement. Its precisely delineated splits, standardized evaluation, and well-studied augmentation pipeline collectively support fair, transparent, and iterative scientific progress in the field.