MambaRate: Multi-Rate MOS Prediction

• MambaRate is a speech quality assessment model designed to predict MOS across multiple sampling rates (16, 24, 48 kHz) with minimized bias.
• It combines standardized preprocessing, SSL feature extraction, and selective state space modeling to ensure robust cross-rate generalization.
• The model employs Gaussian RBF encoding of MOS targets and MSE loss, achieving high correlation and improved error metrics in MOS prediction.

Searching arXiv for the named paper and closely related uses of “MambaRate” to ground the article. arxiv_search(query="(Kakoulidis et al., 16 Jul 2025) MambaRate Speech Quality Assessment Across Different Sampling Rates", max_results=5) arxiv_search({"query":"(Kakoulidis et al., 16 Jul 2025) MambaRate Speech Quality Assessment Across Different Sampling Rates","max_results":5}) MambaRate is a speech quality assessment model for non-intrusive prediction of Mean Opinion Scores (MOS) across different waveform sampling rates. It was proposed for AudioMOS Challenge 2025 Track 3, which targets MOS prediction for speech at 16 kHz, 24 kHz, and 48 kHz, and it is designed to predict MOS with limited bias regarding sampling rate by combining standardized preprocessing, self-supervised speech representations, selective state space modeling, and Gaussian radial basis function target encoding (Kakoulidis et al., 16 Jul 2025).

1. Research setting and objective

MambaRate addresses a failure mode of conventional non-intrusive MOS predictors: most are trained and validated on 16 kHz datasets, which reduces their robustness when deployed on high-fidelity studio speech such as 48 kHz material or mixed-bandwidth audio. When sampling rates vary, both the spectral content and the acoustic distortions perceived by human listeners change, which can bias models that learn sampling-rate-specific cues. The stated objective is therefore to predict MOS with limited sampling-rate bias, without requiring metadata such as system or listener IDs (Kakoulidis et al., 16 Jul 2025).

AudioMOS Challenge 2025 Track 3 formalizes this setting. Its training data contain 400 audio files at 16 kHz, 24 kHz, or 48 kHz, with MOS ratings from 10 listeners across 3 listening tests, where each test contains waveforms of a single sampling rate. Systems are evaluated on unseen audio across all three rates, and the official ranking is based on system-level Spearman’s Rank Correlation Coefficient (SRCC). The challenge thus emphasizes cross-rate generalization rather than performance in a conventional single-rate regime.

Within that setting, MambaRate is defined by four operational choices: SV56 amplitude normalization, resampling all inputs to a common 16 kHz rate for SSL feature compatibility, extraction of robust SSL embeddings offline, and training on rate-aggregated targets such as the median or mean across listening tests of different sampling rates. The stated rationale is that such targets discourage rate-specific label leakage.

2. Signal processing pipeline and model architecture

The input pipeline begins with SV56 amplitude normalization, after which all challenge and BVCC audios are downsampled to 16 kHz for compatibility with the SSL models. The initial submission system, denoted T16, uses WavLM-Large embeddings extracted offline from publicly available weights and specifically uses the 6th layer embeddings. Later ablations examine layers 7–10 and the final layer. Additional representation families are also evaluated: Whisper large-v3-turbo final-layer features and VGGish final-layer features, both from public checkpoints (Kakoulidis et al., 16 Jul 2025).

MambaRate does not operate directly on raw waveforms. It takes per-frame SSL embeddings and reduces them through a learned downsampling head to a 64-channel sequence. That front end consists of a 1D convolution with output channels $= 64$, kernel size $= 3$, and stride $= 1$, followed by layer normalization and Mish activation. The exact native dimensionality and frame rate of the SSL backbones are not required by the downstream model.

The sequence model uses Selective State Space Modeling with Mamba-2. Each block comprises a Mamba-2 layer with $d_{\text{model}} = 64$, $d_{\text{state}} = 32$, $d_{\text{conv}} = 4$, and $\text{expand} = 8$, followed by a feedforward layer with Mish activation. The generic continuous-time and discrete-time state-space formulations are given as

$$x'(t) = A x(t) + B u(t), \qquad y(t) = C x(t) + D u(t),$$

and

$$s_t = A s_{t-1} + B u_t, \qquad y_t = C s_t + D u_t.$$

In Mamba-2, the selectivity stems from data-dependent parameterization and efficient scanning over sequences, enabling linear-time/state complexity in sequence length. The convolutional component $d_{\text{conv}}$ and expansion factor $= 3$0 provide local mixing and channel-wise expansion before gating.

The utterance-level prediction head is a multi-perceptron block with two linear layers. Average pooling along the sequence length is applied before the last linear layer, a sigmoid activation is applied on outputs, and the output dimensionality is 16.

3. Target representation, decoding, and loss

A distinctive feature of MambaRate is its Gaussian radial basis function encoding of scalar MOS targets. For centers $= 3$1 evenly spaced in $= 3$2 with $= 3$3, an utterance MOS $= 3$4 is encoded as

$= 3$5

The second center is given as $= 3$6, with $= 3$7, $= 3$8, and $= 3$9. Before encoding, small noise $= 1$0 is added to $= 1$1 as a robustness trick (Kakoulidis et al., 16 Jul 2025).

The model is trained to regress the 16-dimensional RBF codes with mean squared error:

$= 1$2

For reference, the scalar MOS objective is

$= 1$3

In practice, training is performed on the RBF vectors rather than directly on scalar MOS.

Decoding back to a scalar MOS is performed by argmax over the predicted 16-dimensional RBF vector:

$= 1$4

Evaluation then uses the scalar MOS obtained from the selected center. This representation replaces direct scalar regression with a structured continuous code over the MOS interval $= 1$5.

4. Data, training protocol, and evaluation criteria

Two datasets are central to the reported experiments. The AudioMOS Challenge 2025 Track 3 training set provides the multi-rate benchmark, while BVCC serves as a public single-rate dataset for benchmarking and pre-training. BVCC contains listening tests at 16 kHz and has established splits and public baseline models. The challenge data provide no metadata on the systems or tasks, such as voice conversion, that produced the audios (Kakoulidis et al., 16 Jul 2025).

For the T16 submission, the challenge training data are split 90\%/10\% into train and validation subsets. For later ablations, a 70\%/15\%/15\% train/validation/test split is used. Early experiments, including T16, use the median of ratings per utterance across listening tests as targets. Later ablations use the mean of ratings per utterance, and the reported outcome is that mean targets improve performance.

Optimization uses AdamW with learning rate $= 1$6 and weight decay $= 1$7, a cosine annealing scheduler with $= 1$8, and early stopping with patience $= 1$9 and $d_{\text{model}} = 64$0. The model is described as lightweight, with approximately 900k parameters, and in the few-shot, no-pretraining configuration it trains in minutes and for fewer than 30 epochs due to early stopping. The submitted T16 system does not pre-train on external data, whereas additional experiments pre-train on BVCC and then fine-tune on challenge data.

Reported evaluation metrics are MSE, LCC, SRCC, and KTAU, with the challenge ranking systems by system-level SRCC. The paper gives the standard formulas for system-level Spearman SRCC,

$d_{\text{model}} = 64$1

for Pearson’s LCC,

$d_{\text{model}} = 64$2

and for MSE,

$d_{\text{model}} = 64$3

5. Empirical results and ablation findings

In the conventional single-rate BVCC setting at 16 kHz, MambaRate improves MSE from 0.277 to 0.209 and KTAU from 0.690 to 0.736 relative to the baseline, while LCC is 0.865 versus 0.869 and SRCC is 0.866 versus 0.869, indicating comparable correlation and improved error and Kendall consistency (Kakoulidis et al., 16 Jul 2025).

In the challenge-related multi-rate setting, where ratings are averaged across all sampling rates per utterance, the comparison is more pronounced. The baseline used zero-shot on BVCC yields MSE 1.069, LCC 0.748, SRCC 0.522, and KTAU 0.350. A fine-tuned baseline using BVCC plus challenge data yields MSE 0.336, LCC 0.854, SRCC 0.853, and KTAU 0.686. MambaRate trained only on challenge data yields MSE 0.086, LCC 0.871, SRCC 0.879, and KTAU 0.738. When pre-trained on BVCC and fine-tuned on challenge data, MambaRate reaches MSE 0.071, LCC 0.891, SRCC 0.930, and KTAU 0.822. The paper explicitly notes that MambaRate trained only on challenge data outperforms both the baseline zero-shot and the fine-tuned baseline, and that BVCC pre-training further improves the model in the multi-rate setting.

The official AudioMOS Track 3 outcome reflects a stricter few-shot condition. The initial MambaRate submission, T16, outperformed the pre-trained baseline B03 by approximately 14\% and ranked fourth out of five, approximately 6\% behind the winning system. This official ranking is based on system-level SRCC.

Ablations indicate that the initial T16 configuration is not the strongest variant. For WavLM-Large layer-wise experiments with mean targets, Layer 6 yields utterance-level metrics of MSE 0.057, LCC 0.922, SRCC 0.899, and KTAU 0.782, and system-level metrics of MSE 0.040, LCC 0.963, SRCC 0.976, and KTAU 0.929. Layer 7 reaches system-level MSE 0.017, LCC 0.982, SRCC 0.976, and KTAU 0.928. Layer 9 yields utterance-level MSE 0.038, LCC 0.937, SRCC 0.911, and KTAU 0.801, and system-level MSE 0.033, LCC 0.967, SRCC 0.976, and KTAU 0.928. The paper summarizes this as an indication that mid-to-upper WavLM layers, especially 7–9, yield strong system-level SRCC and can surpass the initial layer-6 T16 configuration.

Cross-model ablations show complementary behavior across representation families. With mean targets, WavLM final-layer features give utterance-level SRCC 0.897 and KTAU 0.770, Whisper large-v3-turbo final features give utterance-level SRCC 0.925 and KTAU 0.816, and VGGish final features give system-level SRCC 0.952 and KTAU 0.857. The reported interpretation is that Whisper embeddings improve utterance-level SRCC and KTAU, while VGGish yields strong system-level SRCC and KTAU. These ablations outperform the initial T16 setup.

6. Bias mitigation, efficiency, limitations, and future directions

The paper presents MambaRate as a model for handling sampling-rate bias rather than merely tolerating mixed-rate input. Four strategies are emphasized. First, multi-rate training uses data containing 16, 24, and 48 kHz utterances. Second, targets aggregate ratings across listening tests, discouraging learning of rate-specific biases. Third, all inputs are downsampled to 16 kHz and amplitude-normalized with SV56, which standardizes the representation space for the SSL encoders and yields scale-invariant, sampling-rate-agnostic input features. Fourth, robust SSL embeddings and selective SSM dynamics are used to focus on content-related quality cues rather than sampling-rate artifacts (Kakoulidis et al., 16 Jul 2025).

The evidence presented for reduced bias is empirical. In the three-rate challenge scenario, MambaRate trained only on challenge data achieves utterance-level SRCC 0.879, exceeding the zero-shot baseline SRCC 0.522 and the fine-tuned baseline SRCC 0.853. With BVCC pre-training, it improves further to SRCC 0.930. The official Track 3 result, where the few-shot T16 system without pre-training exceeds the pre-trained baseline B03 by approximately 14\%, is reported as consistent with reduced rate-driven performance gaps.

The implementation is intentionally small. The parameter count is approximately 900k, the model trains in minutes, and inference uses linear-time scanning with Mamba-2 SSM and small heads. Support for 48 kHz audio is obtained by resampling to 16 kHz, and SSL features are extracted offline from public models, which keeps the online network compact. Reproducibility is supported by the use of public checkpoints, standardized preprocessing, and explicit hyperparameters.

The limitations are also explicit. Resampling to 16 kHz may discard information present in 24 kHz and 48 kHz recordings, so some high-band artifacts relevant to MOS may be under-modeled. The challenge dataset lacks system and task metadata, and broader validation is still needed for unseen synthesis or enhancement systems, languages, and noise conditions. Sensitivity to intermediate or higher sampling rates is not explicitly tested. Suggested future directions include combining the best-performing SSL layers, evaluating on additional public datasets such as SOMOS and VCC2018, augmenting data by upsampling or bandwidth extension followed by new listening tests, and exploring calibration techniques, uncertainty modeling, and multi-rate augmentation.

7. Terminological ambiguity and related uses

Within arXiv, “MambaRate” most directly denotes the speech MOS predictor introduced in “MambaRate: Speech Quality Assessment Across Different Sampling Rates” (Kakoulidis et al., 16 Jul 2025). The same string, however, is used in distinct ways in later or adjacent literature, and these usages are not interchangeable.

In “Timesteps of Mamba Align with Human Reading Times,” “MambaRate” refers to a per-word quantity $d_{\text{model}} = 64$4 derived from Mamba’s input-conditioned discretization timestep $d_{\text{model}} = 64$5. It is computed by aggregating the per-token $d_{\text{model}} = 64$6 vector across channels and then mapping subword tokens to words; in that work, the resulting per-word signal is analyzed as a predictor of human reading times (Yamamoto et al., 29 Jun 2026).

In “A Mamba-based Perceptual Loss Function for Learning-based UGC Transcoding,” “MambaRate” denotes leveraging a Mamba-based perceptual quality model within the rate–distortion optimization of learning-based UGC transcoding. There it refers to a perceptual loss term $d_{\text{model}} = 64$7 that augments or replaces pixel-based distortion during codec optimization (Qi et al., 26 Mar 2026).

In “A Novel Mamba-based Sequential Recommendation Method,” the authors describe Hydra as a concrete instantiation of a “MambaRate” engine for efficient sequence-based rating or prediction, built from multi-head latent Mamba layers and a final softmax prediction head (Yuan, 10 Apr 2025). By contrast, the WebRTC adaptive bitrate system “Mamba: Bringing Multi-Dimensional ABR to WebRTC” explicitly does not use the term “MambaRate”; its algorithm is simply named Mamba (Li et al., 2023).

These multiple usages indicate that “MambaRate” has become a label applied to several Mamba-based prediction or scoring constructs. In current arXiv usage, however, the title-bearing and primary referent remains the speech quality assessment model for cross-sampling-rate MOS prediction (Kakoulidis et al., 16 Jul 2025).