SquimSubjective

class torchaudio.models.SquimSubjective(ssl_model: Module, projector: Module, predictor: Module)

Speech Quality and Intelligibility Measures (SQUIM) model that predicts subjective metric scores for speech enhancement (e.g., Mean Opinion Score (MOS)). The model is adopted from NORESQA-MOS [Manocha and Kumar, 2022] which predicts MOS scores given the input speech and a non-matching reference.

Parameters:

• ssl_model (torch.nn.Module) – The self-supervised learning model for feature extraction.

• projector (torch.nn.Module) – Projection layer that projects SSL feature to a lower dimension.

• predictor (torch.nn.Module) – Predict the subjective scores.

Tutorials using SquimSubjective:

Torchaudio-Squim: Non-intrusive Speech Assessment in TorchAudio

Torchaudio-Squim: Non-intrusive Speech Assessment in TorchAudio

Methods

forward

SquimSubjective.forward(waveform: Tensor, reference: Tensor)

Predict subjective evaluation metric score.

Parameters:

• waveform (torch.Tensor) – Input waveform for evaluation. Tensor with dimensions (batch, time).

• reference (torch.Tensor) – Non-matching clean reference. Tensor with dimensions (batch, time_ref).

Returns:

Subjective metric score. Tensor with dimensions (batch,).

Return type:

(torch.Tensor)

Factory Functions

squim_subjective_model	Build a custome torchaudio.prototype.models.SquimSubjective model.
squim_subjective_base	Build torchaudio.prototype.models.SquimSubjective model with default arguments.