Implementation:Facebookresearch Audiocraft ViSQOL Wrapper

Knowledge Sources	Facebookresearch_Audiocraft
Domains	Audio_Metrics, Speech_Quality
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for computing ViSQOL (Virtual Speech Quality Objective Listener) perceptual quality scores by wrapping Google's ViSQOL binary.

Description

ViSQOL provides a Python wrapper around the external ViSQOL binary. It handles audio resampling to the required sample rate (48 kHz for audio mode, 16 kHz for speech mode), temporary file management, subprocess execution, and result parsing. It supports both audio quality assessment and speech quality assessment modes.

Usage

Import this class when computing perceptual audio quality metrics for compression model evaluation. Requires the ViSQOL binary to be installed.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/metrics/visqol.py
• Lines: 1-216

Signature

class ViSQOL:
    def __init__(self, bin: tp.Union[Path, str], mode: str = "audio",
                 model: str = "libsvm_nu_svr_model.txt", debug: bool = False): ...
    def __call__(self, ref_sig: torch.Tensor, deg_sig: torch.Tensor, sr: int,
                 pad_with_silence: bool = False) -> float: ...

Import

from audiocraft.metrics.visqol import ViSQOL

I/O Contract

Inputs

Name	Type	Required	Description
ref_sig	torch.Tensor	Yes	Reference audio [B, C, T]
deg_sig	torch.Tensor	Yes	Degraded/generated audio [B, C, T]
sr	int	Yes	Sample rate of input audio

Outputs

Name	Type	Description
score	float	Average MOSLQO quality score

Related Pages

• Principle:Facebookresearch_Audiocraft_Compression_Quality_Evaluation