DNSMOS-C: End-to-End Speech Quality Metric

• The paper extends DNSMOS Pro by integrating a MOS-guided triplet-based contrastive loss to reorganize the latent space for improved MOS alignment.
• It introduces a unified framework that couples a convolutional encoder with Gaussian regression to jointly learn speech representations and MOS prediction.
• It maintains computational efficiency and simplicity, making it ideal for non-intrusive evaluation of noise suppression in real-world recordings.

DNSMOS-C is a compact end-to-end speech quality assessment model that extends the DNSMOS Pro framework by integrating a MOS-guided triplet-based contrastive loss. It belongs to the class of non-intrusive perceptual objective speech quality metrics, a setting in which human subjective evaluation remains the gold standard, perceptual objective metrics serve as a proxy for subjective scores, and clean reference signals are often unavailable in real recordings. Within the DNSMOS lineage, the earlier DNSMOS work introduced a multi-stage self-teaching based perceptual objective metric designed to evaluate noise suppressors, whereas DNSMOS-C preserves a single, unified end-to-end framework while explicitly reorganizing latent representations with respect to perceptual quality (Reddy et al., 2020, Liang et al., 25 Jun 2026).

1. Position within no-reference speech quality assessment

Conventional and widely used perceptual objective metrics require a reference clean speech signal, which is unavailable in real recordings. The no-reference approaches correlate poorly with human ratings and are not widely adopted in the research community. One of the biggest use cases of such perceptual objective metrics is the evaluation of noise suppression algorithms (Reddy et al., 2020).

DNSMOS-C is situated in this no-reference setting but departs from earlier end-to-end formulations by adding explicit MOS-guided contrastive supervision to the intermediate embedding space. It is described as a compact end-to-end speech quality assessment model that extends the DNSMOS Pro framework by integrating a MOS-guided triplet-based contrastive loss. Unlike prior methods that depend on large pre-trained self-supervised learning encoders and multi-stage training, DNSMOS-C jointly learns speech representations and MOS regression within a single, unified framework (Liang et al., 25 Jun 2026).

A central distinction in the DNSMOS-C formulation is therefore not merely architectural compactness, but the imposition of a quality-aware geometry on the latent space. The model is designed to improve correlation metrics over DNSMOS Pro and to achieve better generalization on challenging out-of-domain test sets, while preserving the simplicity and efficiency associated with DNSMOS Pro (Liang et al., 25 Jun 2026).

2. Signal pipeline and model architecture

The input pipeline is fixed and explicit. Audio clips are first down-sampled to 16 kHz, then padded or cropped to exactly 10 s. A 20 ms Hann window with 10 ms hop is applied to compute a log-magnitude spectrogram, and values are clipped to $[-7, 7]$. This spectrogram, organized as time by frequency, is fed directly into a small convolutional encoder (Liang et al., 25 Jun 2026).

DNSMOS-C borrows the four-layer convolution to global max-pooling encoder from DNSMOS Pro, producing a $D=64$ embedding $e \in \mathbb{R}^{64}$ for each clip. Conv1 through Conv4 each apply 2D convolutions over time and frequency, with ReLU activations and optional batch normalization, halving the time axis and doubling channels at each stage, for example 32 to 64 to 128 to 256 filters. Global max-pooling across the time and frequency axes collapses the representation to a 64-dimensional vector (Liang et al., 25 Jun 2026).

A lightweight prediction head then maps the embedding to the parameters of a Gaussian posterior over the MOS. The head consists of three fully connected layers with ReLU, for example 64 to 64 to 64 to 64, followed by two parallel linear outputs: one for the mean $\mu(x)$ and one for the log-variance $\log \sigma^2(x)$. At inference, $\hat y = \mu(x)$ is reported as the predicted MOS, while $\sigma^2(x)$ can serve as an uncertainty estimate (Liang et al., 25 Jun 2026).

This design couples compact feature extraction with probabilistic regression. The architecture remains small, but the prediction target is not restricted to a point estimate alone; it also models output variance through the Gaussian parameterization.

3. Contrastive objective and regression formulation

DNSMOS-C augments the usual MOS regression loss with a triplet-style contrastive term that explicitly pulls together clips of similar MOS and pushes apart those of dissimilar MOS. Triplets $(i,j,k)$ are formed so that sample $i$ is the anchor, $j$ has a MOS close to $D=64$0's as the positive, and $D=64$1 has a substantially different MOS as the negative. In practice, a small threshold $D=64$2 is chosen, for example $D=64$3 MOS, to select $D=64$4 and $D=64$5 (Liang et al., 25 Jun 2026).

The contrastive loss over encoder embeddings is given by

$D=64$6

where $D=64$7 is the Euclidean distance in the 64-dimensional latent space, and $D=64$8 is the margin. In the reported experiments, $D=64$9 was set to $e \in \mathbb{R}^{64}$0 (Liang et al., 25 Jun 2026).

The regression component is implemented as a Gaussian negative log-likelihood,

$e \in \mathbb{R}^{64}$1

which is described as fully equivalent to learning both mean and variance. For clarity, the formulation may also be viewed through the point-estimate loss

$e \in \mathbb{R}^{64}$2

The two objectives are combined in a single joint optimization:

$e \in \mathbb{R}^{64}$3

In all experiments, $e \in \mathbb{R}^{64}$4 was found to yield the best trade-off via small grid searches on each validation set (Liang et al., 25 Jun 2026).

The conceptual role of this formulation is explicit: regression preserves direct MOS supervision, while the triplet term regularizes representation geometry so that perceptually similar clips are close in the embedding space. A plausible implication is that the model is encouraged to encode a more monotonic notion of quality, rather than organizing latent structure primarily by distortion identity.

4. Training protocol and datasets

DNSMOS-C is trained as three independent variants on BVCC, Tencent, and NISQA_SIM. All splits follow the original DNSMOS Pro paper (Liang et al., 25 Jun 2026).

Dataset	Split and language	Data characteristic
BVCC	$e \in \mathbb{R}^{64}$5k / $e \in \mathbb{R}^{64}$6k / $e \in \mathbb{R}^{64}$7k; English	MOS from TTS/VC systems
Tencent	$e \in \mathbb{R}^{64}$8k / $e \in \mathbb{R}^{64}$9k / $\mu(x)$0k; Mandarin	synthetic distortions
NISQA_SIM	$\mu(x)$1k / $\mu(x)$2k / $\mu(x)$3k; English	simulated distortions

The optimizer is Adam with $\mu(x)$4 and $\mu(x)$5, and the learning rate is $\mu(x)$6. Training uses batch size 32 triplets per batch, organized as anchor, positive, and negative. The reported schedule is 500 full-dataset passes, with model selection by highest LCC on the validation set. The contrastive margin is $\mu(x)$7, and the contrastive weight is $\mu(x)$8 (Liang et al., 25 Jun 2026).

A methodological point emphasized in the description is the contrast between DNSMOS-C and multi-stage pipelines. Unlike prior work, specifically SCOREQ, which first pre-trains an encoder with contrastive loss and then freezes it to train a separate MOS head, DNSMOS-C jointly learns both objectives end-to-end in a single stage, preserving DNSMOS Pro’s simplicity and speed (Liang et al., 25 Jun 2026).

5. Empirical performance and generalization behavior

The reported evaluation metrics are mean squared error (MSE), Pearson correlation (LCC), and Spearman rank correlation (SRCC). Numbers are mean $\mu(x)$9 standard deviation over 10 runs with different random seeds (Liang et al., 25 Jun 2026).

For in-domain test sets, DNSMOS-C consistently raises correlation metrics over DNSMOS Pro:

Dataset	DNSMOS Pro	DNSMOS-C
BVCC	MSE $\log \sigma^2(x)$0, LCC $\log \sigma^2(x)$1, SRCC $\log \sigma^2(x)$2	MSE $\log \sigma^2(x)$3, LCC $\log \sigma^2(x)$4, SRCC $\log \sigma^2(x)$5
NISQA_SIM	MSE $\log \sigma^2(x)$6, LCC $\log \sigma^2(x)$7, SRCC $\log \sigma^2(x)$8	MSE $\log \sigma^2(x)$9, LCC $\hat y = \mu(x)$0, SRCC $\hat y = \mu(x)$1
Tencent	MSE $\hat y = \mu(x)$2, LCC $\hat y = \mu(x)$3, SRCC $\hat y = \mu(x)$4	MSE $\hat y = \mu(x)$5, LCC $\hat y = \mu(x)$6, SRCC $\hat y = \mu(x)$7

Across all three training domains, DNSMOS-C is reported to consistently raise correlation, while matching or slightly improving MSE. The numerical pattern also shows an important nuance: the improvements are most consistent in LCC and SRCC, whereas MSE is not uniformly reduced, as illustrated by the NISQA_SIM in-domain result (Liang et al., 25 Jun 2026).

Out-of-domain generalization is reported for models trained on NISQA_SIM and tested on three unseen NISQA splits. On NISQA_TEST_FOR, LCC rises from $\hat y = \mu(x)$8 to $\hat y = \mu(x)$9 and SRCC from $\sigma^2(x)$0 to $\sigma^2(x)$1, with a small regression drift in MSE from $\sigma^2(x)$2 to $\sigma^2(x)$3. On NISQA_TEST_P501, LCC increases from $\sigma^2(x)$4 to $\sigma^2(x)$5 and SRCC from $\sigma^2(x)$6 to $\sigma^2(x)$7, while MSE changes from $\sigma^2(x)$8 to $\sigma^2(x)$9. On NISQA_TEST_LIVETALK, LCC and SRCC are near-identical, with a slight MSE increase (Liang et al., 25 Jun 2026).

These results delimit the model’s empirical contribution precisely. DNSMOS-C improves rank-order consistency and linear correlation more reliably than it improves absolute squared error, especially under domain shift. This makes correlation-oriented reading of the results essential.

6. Latent-space organization, interpretability, and computational profile

Two complementary analyses are reported to confirm that the triplet loss reshapes the 64-dimensional embedding into a smoother “quality manifold” (Liang et al., 25 Jun 2026).

The first is PCA-based quality correlation on TCD-VoIP, described as unseen distortions with 5 classes. Embeddings are projected onto their first two principal components, and the multiple correlation coefficient $(i,j,k)$0 between $(i,j,k)$1 and true MOS is computed. DNSMOS Pro attains $(i,j,k)$2 with LCC on predicted MOS $(i,j,k)$3, whereas DNSMOS-C reaches $(i,j,k)$4 with LCC $(i,j,k)$5. The visual interpretation provided is that DNSMOS-C’s PC1 aligns strongly with the MOS color gradient and distortion clusters collapse into a continuous spectrum, whereas DNSMOS Pro shows tight clusters per distortion type with less smooth MOS progression (Liang et al., 25 Jun 2026).

The second analysis concerns clustering of impairment and noise types. On LibriAugmented1600 with 16 impairment types, the encoder of DNSMOS Pro clusters distortions more distinctly, with accuracy approximately $(i,j,k)$6, than DNSMOS-C, approximately $(i,j,k)$7, which is described as reflecting that contrastive loss trades some distortion-type separability to focus on MOS. On ESC50 with 50 noise categories, DNSMOS-C improves noise-type clustering from $(i,j,k)$8 to $(i,j,k)$9, with the explanation that noise correlates with perceptual quality and is thus reinforced by MOS contrastive learning (Liang et al., 25 Jun 2026).

The computational profile is intentionally unchanged. DNSMOS-C uses the identical encoder and head architecture and the same 64-dimensional pooling as DNSMOS Pro, approximately 500 K parameters total, so there is no additional model-size or inference latency overhead. In deployment, including VoIP and streaming, CPU/GPU throughput and real-time factor remain unchanged. Contrastive sampling and loss are only computed during training; at inference DNSMOS-C is byte-for-byte the same as DNSMOS Pro (Liang et al., 25 Jun 2026).

A common misunderstanding would be to interpret DNSMOS-C as a heavier or staged alternative to DNSMOS Pro. The reported formulation states the opposite: the model keeps the same inference graph and runtime characteristics, and the difference lies in the training objective and the resulting latent organization. Another misunderstanding would be to read the latent-space analyses as demonstrating uniformly better separability for all nuisance factors; the LibriAugmented1600 result explicitly shows a trade-off in which some impairment-type clustering is reduced while MOS alignment is strengthened (Liang et al., 25 Jun 2026).