DNSMOS-Filtered Separated Vocals

• DNSMOS-filtered separated vocals are produced by combining advanced source separation algorithms with DNSMOS evaluation to enhance quality and suppress artifacts.
• They improve singing voice conversion, ASR, and neural vocoder training by ensuring natural timbre and reduced residual noise.
• DNSMOS integration enables dynamic quality control and robust performance assessment in both research and practical audio production.

DNSMOS-filtered separated vocals refers to singing voice recordings that have been extracted from polyphonic mixtures using source separation algorithms and then further filtered or evaluated using the DNSMOS (Deep Noise Suppression Mean Opinion Score) metric or methodology. This combined approach synthesizes advances in both robust vocal separation and perceptual speech quality estimation, enabling improved noise-robust, artifact-suppressed, and more natural-sounding vocal tracks for downstream applications in singing voice conversion, music production, ASR for singing, and related tasks.

1. Definition and Overview

DNSMOS is a deep-learning-based perceptual objective metric designed to evaluate noise suppressors and filter enhanced speech with a high correlation to human-subjective mean opinion scores (Reddy et al., 2020, Reddy et al., 2021). Separated vocals refer to the output of source separation models that isolate the singing voice from its musical accompaniment or environmental noise. DNSMOS-filtered separated vocals are produced by applying such separation algorithms and then leveraging DNSMOS—either as a post-filtering stage or as an objective for quality control—to select, enhance, or further train on vocal signals that score highly in terms of human perceptual quality.

This concept underpins recent efforts in singing voice conversion (SVC), robust ASR for mixed music scenarios, neural vocoder training for dry vocals, and artifact suppression for real-world deployment, as seen in frameworks such as R2-SVC (Zheng et al., 23 Oct 2025), JRSV (Bai et al., 17 Apr 2024), and neural vocoder feature-based SVS (Im et al., 2022).

2. Source Separation and Artifact Suppression

Separation of vocals from accompaniment in polyphonic music relies on a diverse suite of algorithms, including informed group-sparse representations (Chan et al., 2018), deep convolutional neural networks trained on ideal binary masks (Lin et al., 2018), multichannel DL+CNMF systems (Muñoz-Montoro et al., 2020), self-attention convolutional networks (Liu et al., 2020), and UNet-derived SVS for downstream singing voice detection (Sun et al., 2020).

These algorithms produce vocal tracks that may contain residual noise, reverberation, or artifacts introduced by imperfect modeling, mask leakage, or musical interference. Artifact suppression, dereverberation, and quality consistency remain central challenges. To address these, post-processing based on perceptual metrics enables systems to filter out outputs likely to degrade intelligibility or naturalness, as assessed by objective measures such as DNSMOS.

In advanced singing voice conversion systems, separated vocals are simulated to include realistic noise and artifact profiles, fitting the scenarios encountered in real-world deployment (Zheng et al., 23 Oct 2025). This simulation-based robustness enhancement applies random F₀ perturbations (jitter, glide, jump) and mixes in reverberation or echo to augment the diversity and realism of the training data.

3. DNSMOS Metric and Filtering Workflow

DNSMOS predicts perceptual speech or vocal quality from enhanced signals without the need for a clean reference, supporting non-intrusive real-time evaluation (Reddy et al., 2020, Reddy et al., 2021). Formally, the DNSMOS score for a given signal $x$ is

$\text{DNSMOS}(x) = f(x)$

where $f(\cdot)$ is a DNN trained on a large corpus with subjective MOS labels. Differential DNSMOS (DMOS) can be used for quality improvement quantification:

$$\text{DMOS} = \text{MOS}_{\text{enhanced}} - \text{MOS}_{\text{noisy}}$$

Filtering separated vocals involves scoring all candidate tracks and either selecting those above a threshold, passing the signal through a DNSMOS-tuned post-filter (to suppress artifacts), or using DNSMOS as an auxiliary loss function during model training (to directly optimize for perceptual quality).

DNSMOS filtering has been utilized for model adaptation and training data selection in zero-shot SVC (Zheng et al., 23 Oct 2025), as well as for perceptual quality control in ASR pipelines that separate and transcribe singing voice and speech in overlapping scenarios (Bai et al., 17 Apr 2024).

4. Integration with Robust Singing Applications

Modern robust SVC systems such as R2-SVC (Zheng et al., 23 Oct 2025) enrich their singing-specific timbre and style extractor (SETSE) by integrating DNSMOS-filtered separated vocals into their domain data. This practice exposes the learning model to a realistic distribution of vocal timbres and expressive styles—including those affected by music separation artifacts—while maintaining high perceptual quality, as ensured by DNSMOS evaluation.

The model architecture typically incorporates simulation-based robustness enhancement for both input (simulated F₀ perturbations, artifact mixing) and training data (DNSMOS-filtered separated vocals, public singing corpora), and utilizes the Neural Source-Filter (NSF) model to explicitly represent harmonic excitation and noise components, thus improving both naturalness and controllability in the converted vocal output.

R2-SVC formalizes its conversion process as follows:

• Robust transformed input: $$x^{(\text{aug}, F_0^{(\text{pert})})} = \mathcal{R}(x, F_0)$$
• Singing-specific embedding (incl. DNSMOS-filtered vocals): $s = \mathcal{E}(x; \mathcal{D}_{\text{sing}})$
• NSF feature extraction: $h_{\text{nsf}} = \mathcal{N}(F_0, t)$
• Output generation: $$\hat{y} = \mathcal{F}_\theta(x^{(\text{aug}, F_0^{(\text{pert})})}, s, h_{\text{nsf}})$$

5. DNSMOS-Filtered Vocals in Neural Vocoder Training and Evaluation

Neural-vocoder-based singing voice separation (Im et al., 2022) employs mel-spectrogram features, which are either directly predicted or obtained via binary mask application on the mixed track. DNSMOS filtering is well-suited to this architecture, as both the model's output and the perceptual evaluation operate in the waveform domain.

A singing voice detector is integrated to mask segments where no vocal activity occurs, reducing false positives and further enabling DNSMOS-driven selection and enhancement of usable vocal segments. Objective metrics such as SiSPNR (scale-invariant spectrogram-to-noise ratio) and SPDR (spectrogram to distortion ratio) complement DNSMOS scores in evaluating dereverberation and separation quality.

DNSMOS filtering within this context refines pre-vocoded signals, ensures that only perceptually high-quality segments are synthesized and used, and can be integrated as an additional training objective to maximize naturalness and clarity.

6. Multimodal and Lightweight Systems: DNSMOS Scoring, Quality Control

Lightweight source separation frameworks such as DTTNet (Chen et al., 2023), as well as multimodal (e.g., audiovisual separation (Li et al., 2021)) systems, have incorporated or suggested DNSMOS assessment for quality control. These systems typically report separation quality via chunk-level SDR, but DNSMOS provides a complementary evaluation in terms of timbre preservation, intelligibility, and absence of residual musical noise.

DNSMOS scoring is also relevant for real-time and embedded separation applications, as it can be computed rapidly for dynamic parameter tuning or segment rejection.

7. Experimental Outcomes and Future Directions

Empirical results across studies indicate that DNSMOS-filtered separated vocals, when used for model selection, training, and post-filtering, yield improved speaker similarity (SPK-SIM), maintain competitive intelligibility (low CER), and achieve high aesthetic scores in the context of singing voice conversion and robust ASR (Zheng et al., 23 Oct 2025, Bai et al., 17 Apr 2024). The use of DNSMOS also facilitates adaptation to challenging industrial noise conditions and artifact profiles, with enhanced perceptual clarity and naturalness over baselines.

Future directions include:

• Joint optimization of separation and DNSMOS (or similar perceptual loss), potentially via reference-free mediation networks (e.g., PESQNet (Xu et al., 2021)).
• Expansion of DNSMOS to music-domain subjective criteria, explicitly tuning for singing-specific perceptual attributes.
• Real-time DNSMOS filtering for dynamic, adaptive separation workflows.
• Integration into multimodal systems (audio-visual, lyric-to-vocal mapping), where perceptual evaluation is critical for downstream utility.

Table: Approaches and DNSMOS Integration

Separation Approach	DNSMOS Role	Output/Metric
R2-SVC (SETSE, NSF) (Zheng et al., 23 Oct 2025)	Data selection, training, filtering	SPK-SIM, CER, DNSMOS
Neural-vocoder SVS (Im et al., 2022)	Post-filtering, objective scoring	SiSPNR, SPDR, DNSMOS
Lightweight DTTNet (Chen et al., 2023)	Scoring, segment rejection	cSDR, DNSMOS suggest
Joint ASR+SVS (JRSV) (Bai et al., 17 Apr 2024)	Enhancement, quality assurance	CER, SDRi, DNSMOS

The table summarizes methods for separated vocal generation and the manner in which DNSMOS is utilized for filtering, selection, or evaluation, based on documented roles and metrics.

In summary, DNSMOS-filtered separated vocals are produced by combining advanced source separation methods with perceptual quality assessment and filtering, resulting in vocal signals that maintain both technical separation quality and high human-perceived naturalness, timbre, and intelligibility across varied and challenging real-world conditions.