Deep Noise Suppression Mean Opinion Score (DNSMOS)

Module Interface

class torchmetrics.audio.dnsmos.DeepNoiseSuppressionMeanOpinionScore(fs, personalized, device=None, num_threads=None, **kwargs)

Calculate Deep Noise Suppression performance evaluation based on Mean Opinion Score (DNSMOS).

Human subjective evaluation is the ”gold standard” to evaluate speech quality optimized for human perception. Perceptual objective metrics serve as a proxy for subjective scores. The conventional and widely used 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 these perceptual objective metrics is to evaluate noise suppression algorithms. DNSMOS generalizes well in challenging test conditions with a high correlation to human ratings in stack ranking noise suppression methods. More details can be found in DNSMOS paper and DNSMOS P.835 paper.

As input to forward and update the metric accepts the following input

• preds (Tensor): float tensor with shape (...,time)

As output of forward and compute the metric returns the following output

• dnsmos (Tensor): float tensor of DNSMOS values reduced across the batch

  with shape (...,4) indicating [p808_mos, mos_sig, mos_bak, mos_ovr] in the last dim.

Hint

Using this metric requires you to have librosa, onnxruntime and requests installed. Install as pip install torchmetrics['audio'] or alternatively pip install librosa onnxruntime-gpu requests (if you do not have GPU enabled machine install onnxruntime instead of onnxruntime-gpu)

Caution

The forward and compute methods in this class return a reduced DNSMOS value for a batch. To obtain the DNSMOS value for each sample, you may use the functional counterpart in deep_noise_suppression_mean_opinion_score().

Parameters:

• fs (int) – sampling frequency

• personalized (bool) – whether interfering speaker is penalized

• device (Optional[str]) – the device used for calculating DNSMOS, can be cpu or cuda:n, where n is the index of gpu. If None is given, then the device of input is used.

• num_threads (Optional[int]) – number of threads to use for onnxruntime CPU inference.

Raises:

ModuleNotFoundError – If librosa, onnxruntime or requests packages are not installed

Example

>>> from torch import randn
>>> from torchmetrics.audio import DeepNoiseSuppressionMeanOpinionScore
>>> preds = randn(8000)
>>> dnsmos = DeepNoiseSuppressionMeanOpinionScore(8000, False)
>>> dnsmos(preds)
tensor([2.2..., 2.0..., 1.1..., 1.2...], dtype=torch.float64)

plot(val=None, ax=None)

Plot a single or multiple values from the metric.

Parameters:

• val (Union[Tensor, Sequence[Tensor], None]) – Either a single result from calling metric.forward or metric.compute or a list of these results. If no value is provided, will automatically call metric.compute and plot that result.

• ax (Optional[Axes]) – A matplotlib axis object. If provided will add plot to that axis

Return type:

Tuple[Figure, Union[Axes, ndarray]]

Returns:

Figure and Axes object

Raises:

ModuleNotFoundError – If matplotlib is not installed

>>> # Example plotting a single value
>>> import torch
>>> from torchmetrics.audio import DeepNoiseSuppressionMeanOpinionScore
>>> metric = DeepNoiseSuppressionMeanOpinionScore(8000, False)
>>> metric.update(torch.rand(8000))
>>> fig_, ax_ = metric.plot()

>>> # Example plotting multiple values
>>> import torch
>>> from torchmetrics.audio import DeepNoiseSuppressionMeanOpinionScore
>>> metric = DeepNoiseSuppressionMeanOpinionScore(8000, False)
>>> values = [ ]
>>> for _ in range(10):
...     values.append(metric(torch.rand(8000)))
>>> fig_, ax_ = metric.plot(values)

Functional Interface

torchmetrics.functional.audio.dnsmos.deep_noise_suppression_mean_opinion_score(preds, fs, personalized, device=None, num_threads=None)

Calculate Deep Noise Suppression performance evaluation based on Mean Opinion Score (DNSMOS).

Human subjective evaluation is the ”gold standard” to evaluate speech quality optimized for human perception. Perceptual objective metrics serve as a proxy for subjective scores. The conventional and widely used 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 these perceptual objective metrics is to evaluate noise suppression algorithms. DNSMOS generalizes well in challenging test conditions with a high correlation to human ratings in stack ranking noise suppression methods. More details can be found in DNSMOS paper and DNSMOS P.835 paper.

Hint

Using this metric requires you to have librosa, onnxruntime and requests installed. Install as pip install torchmetrics['audio'] or alternatively pip install librosa onnxruntime-gpu requests (if you do not have GPU enabled machine install onnxruntime instead of onnxruntime-gpu)

Parameters:

• preds (Tensor) – […, time]

• fs (int) – sampling frequency

• personalized (bool) – whether interfering speaker is penalized

• device (Optional[str]) – the device used for calculating DNSMOS, can be cpu or cuda:n, where n is the index of gpu. If None is given, then the device of input is used.

• num_threads (Optional[int]) – the number of threads to use for cpu inference. Defaults to None.

Return type:

Tensor

Returns:

Float tensor with shape (...,4) of DNSMOS values per sample, i.e. [p808_mos, mos_sig, mos_bak, mos_ovr]

Raises:

ModuleNotFoundError – If librosa, onnxruntime or requests packages are not installed

Example

>>> from torch import randn
>>> from torchmetrics.functional.audio.dnsmos import deep_noise_suppression_mean_opinion_score
>>> preds = randn(8000)
>>> deep_noise_suppression_mean_opinion_score(preds, 8000, False)
tensor([2.2..., 2.0..., 1.1..., 1.2...], dtype=torch.float64)