---
title: 'GMLv2: Enhanced Generative Machine Listener'
url: https://www.emergentmind.com/topics/gmlv2
type: topic
---

# GMLv2: Enhanced Generative Machine Listener

GMLv2, short for **Enhanced Generative Machine Listener**, is a **reference-based (intrusive) generative model** for predicting subjective audio quality as measured by **MUSHRA scores**. It is designed to estimate not only a point score but a **distribution** of plausible human ratings, thereby providing both expected perceptual quality and uncertainty information. Relative to earlier objective metrics such as PEAQ and ViSQOL, and relative to its predecessor GMLv1, GMLv2 emphasizes bounded score modeling, uncertainty quantification, and broader generalization to both traditional and neural audio codecs [2509.21463].

## 1. Definition and problem setting

GMLv2 addresses the task of predicting subjective listening-test outcomes from a **reference signal** $x$ and a **test (potentially degraded) signal** $\tilde{x}$. Its target variable is the MUSHRA score defined on the interval $[0,100]$, which the model normalizes to $[0,1]$ during training. The model is therefore explicitly **intrusive**: it requires access to both the original signal and the processed signal, rather than operating as a no-reference metric.

A central feature of GMLv2 is that it does not return only a single scalar estimate. Instead, it predicts a **Beta distribution** over normalized listener scores. This design is motivated by three properties stated in the source description: MUSHRA scores are bounded, listener responses exhibit uncertainty, and modern codec distortions are heterogeneous. In that sense, GMLv2 is a probabilistic perceptual metric whose output can be interpreted at the level of both expected quality and confidence.

The model is positioned particularly toward **modern audio coding technologies**, including **neural audio coding (NAC)**. The incorporation of NAC subjective datasets is presented as a key extension beyond GMLv1, intended to improve generalization to newer codec families that are not well covered by older objective measures [2509.21463].

## 2. Signal representation and network architecture

The input to GMLv2 is the pair $(x,\tilde{x})$. For stereo material, both signals are expanded from left-right form into **Left (L), Right (R), Mid**, and **Side** channels, with
$$
M = \frac{L+R}{2}, \qquad S = \frac{L-R}{2}.
$$
This yields a multi-channel representation intended to preserve both channel-specific and inter-channel perceptual structure.

Feature extraction is based on a **Gammatone filterbank**, producing a **Gammatone spectrogram** using only power. The impulse response is given as
$$
h(t) =
\begin{cases}
c\, t^{n-1} \exp(-2\pi b t)\cos(2\pi f_0 t + \phi), & t>0 \\
0, & t<0 .
\end{cases}
$$
The spectrograms for all channels of both the reference and degraded signals are concatenated and used as model input. The source description characterizes this transform as biologically inspired and as mimicking peripheral auditory system filtering.

The backbone is an **Inception-style deep neural network**, following the GMLv1 design. The principal architectural change in GMLv2 occurs at the output layer. Rather than emitting parameters for Gaussian or Logistic likelihoods, the final layer outputs two real values, $\tilde{\alpha}$ and $\tilde{\beta}$, which are transformed into Beta-distribution shape parameters:
$$
\alpha = 1 + \exp(\tilde{\alpha}), \qquad
\beta  = 1 + \exp(\tilde{\beta}).
$$
This parameterization enforces $\alpha,\beta > 1$, which ensures a unimodal distribution on $(0,1)$.

## 3. Beta-distribution formulation

The defining methodological change in GMLv2 is the replacement of GMLv1’s Gaussian/Logistic likelihoods with a **Beta distribution-based loss** [2509.21463]. For normalized score $z \in [0,1]$, the probability density is
$$
g(z;\alpha,\beta) =
\frac{1}{B(\alpha,\beta)} z^{\alpha-1}(1-z)^{\beta-1},
$$
with
$$
B(\alpha,\beta) = \int_0^1 u^{\alpha-1}(1-u)^{\beta-1}\,du.
$$

The reported statistical moments are
$$
\mu = \frac{\alpha}{\alpha+\beta},
$$
$$
\mathrm{Var}[z] = \frac{\alpha\beta}{(\alpha+\beta)^2(\alpha+\beta+1)},
$$
$$
\omega = \frac{\alpha-1}{\alpha+\beta-2} \quad \text{if } \alpha,\beta>1,
$$
and concentration
$$
\kappa = \alpha+\beta.
$$

Training uses the **negative log-likelihood** of the observed normalized MUSHRA score $s$ under the predicted Beta law:
$$
\mathcal{L}(x,\tilde{x},s) = -\log g(s;\alpha,\beta).
$$

The statistical significance of this formulation is threefold. First, the Beta family is **naturally bounded to $[0,1]$**, so it does not assign probability mass outside the feasible score range. Second, it supports **flexible and asymmetric** score distributions, which is relevant when listener ratings are skewed or concentrated near the boundaries. Third, it enables direct uncertainty estimation. The source description states that confidence intervals can be computed by scaling the Beta variance by $100^2$, since MUSHRA is typically reported on $[0,100]$.

A common misconception is to treat GMLv2 as merely a regressor with a probabilistic wrapper. The architecture and loss are instead constructed so that the probability distribution is the primary object of prediction. The mean score is derived from the Beta parameters, and the uncertainty estimate is not post hoc but intrinsic to the output distribution.

## 4. Training data, implementation, and evaluation protocol

GMLv2’s training corpus combines **traditional codecs** and **neural audio coding subjective datasets**. The traditional portion includes material encoded with **AAC, HE-AAC v1/v2, Dolby AC-4**, and related conditions, including stereo and binaural formats at varied bitrates. The NAC portion includes **Encodec, Descript Audio Codec (DAC), MDCTNet, and GAN- or SampleRNN-augmented codecs**, in both mono and stereo. The data summary further notes **binaural and Ambisonic samples via various codecs**, including **Dolby Atmos** and **3GPP IVAS**.

The stated scale is **82,191 total sample pairs**, split into **68,503 traditionals** and **14,688 NACs**, with a **90% training / 10% validation** partition. The description explicitly identifies the inclusion of NAC subjective datasets as a key reason for improved generalization relative to GMLv1.

Training and implementation details are given as follows:

- **Framework**: PyTorch v2.1  
- **Hardware**: 2×NVIDIA A10G GPUs  
- **Batch size**: 8/GPU  
- **Optimizer**: Adam  
- **Learning rate**: $1 \times 10^{-4}$  
- **Training length**: 400k steps  
- **Gammatone spectrogram configuration**: 80 ms window, 20 ms hop, 32 channels, 50 Hz minimum frequency

Evaluation is carried out on the following test sets: **USAC-1/2/3**, **Binaural-1/2**, **NAC Mono**, **NAC Stereo**, and **ODAQ**. The reported objective criteria are **Pearson correlation** $(R_p)$, **Spearman correlation** $(R_s)$, and **Outlier Ratio (OR)**. Pearson correlation is defined as linear fit with ground truth, Spearman correlation as rank-order agreement, and Outlier Ratio as the fraction of cases where prediction diverges significantly from the subjective score [2509.21463].

## 5. Empirical performance

The reported Pearson correlations for GMLv2 are summarized below.

| Test set | GMLv2 $R_p$ |
|---|---:|
| USAC-1 | .92 |
| USAC-2 | .93 |
| USAC-3 | .94 |
| Binaural-1 | .98 |
| Binaural-2 | .99 |
| NAC Mono | .97 |
| NAC Stereo | .95 |
| ODAQ | .83 |
| Average | .94 |

Across these datasets, GMLv2 is reported to deliver the **highest Pearson/Spearman correlations and lowest Outlier Ratios** among the compared methods, including **PEAQ**, **ViSQOL**, **GMLv1**, and **GMLv1\*** [2509.21463]. On average Pearson correlation, the table in the source gives **.53** for PEAQ, **.83** for ViSQOL, **.90** for GMLv1, **.91** for GMLv1\***, and **.94** for GMLv2.

The source description emphasizes performance on **challenging NAC content**. In that regime, the reported $R_p$ values are **.97** for NAC Mono and **.95** for NAC Stereo, compared with **.89** and **.82** for ViSQOL, and **.34** and **.58** for PEAQ. The ODAQ dataset remains comparatively more difficult, but GMLv2 is still reported at **.83**, above the competing values listed in the same table.

A further empirical claim is that GMLv2 reduces the outlier rate by **up to ~47% compared to GMLv1**. The accompanying discussion describes scatter plots in which GMLv2 predictions are tightly clustered along the identity line with minimal outliers, whereas PEAQ and ViSQOL are reported as less reliable in alignment and error distribution.

## 6. Interpretation, scope, and research significance

GMLv2 belongs to a class of perceptual models that treat subjective audio quality prediction as **distribution estimation** rather than scalar regression. In practical terms, it provides a predicted score together with a measure of reliability. The paper therefore frames it as a **scalable and automated framework for perceptual audio quality evaluation**, particularly relevant to research and development in audio coding [2509.21463].

Its principal scope is **codec evaluation**, especially where legacy and neural codecs coexist. Because the training corpus includes both categories, the model is intended to generalize across **diverse content types and codec configurations**. This suggests a role not only in offline benchmarking but also in iterative codec development workflows where repeated listening tests are costly.

Several boundaries of interpretation are equally important. GMLv2 is **reference-based**, so it is not a no-reference quality estimator. It is also trained specifically against **MUSHRA scores**, which means its probabilistic output should be interpreted in the context of that listening-test paradigm. A further point of clarification is that the model’s uncertainty estimate reflects the predicted spread of plausible listener scores under the Beta law; it is not merely an uncertainty of network parameters or ensemble disagreement.

In methodological terms, GMLv2’s distinctive contribution lies in combining four elements within one system: a **reference-based prediction mechanism**, **Gammatone-based auditory front-end**, **Inception-style neural backbone**, and a **Beta distribution-based generative loss**. The resulting model is described as both more accurate and more robust than widely used baselines, particularly once **neural audio coding subjective datasets** are included in training. A plausible implication is that GMLv2 marks a shift from deterministic intrusive metrics toward probabilistic listener-modeling frameworks for perceptual audio assessment.

Source: https://www.emergentmind.com/topics/gmlv2