ViSQOL: definition and a practical example
Short answer: ViSQOL (Virtual Speech Quality Objective Listener) is an open-source algorithm from Google Research that compares received audio with its reference and returns a 1-to-5 MOS-LQO score.
In Portuguese: ViSQOL
What is ViSQOL?
ViSQOL (Virtual Speech Quality Objective Listener) is an algorithm that scores the quality of audio by comparing it with the original. It was created by Google Research, is open source and is published at github.com/google/visqol. The current version is described in the paper “ViSQOL v3” by Chinen et al., presented at the QoMEX conference in 2020.
ViSQOL is a full-reference algorithm: it needs two files, the reference audio and the degraded audio. It aligns the two in time, compares their spectrograms segment by segment and measures the spectro-temporal similarity between them. That similarity is converted into a MOS-LQO score, on the same 1-to-5 scale as MOS.
The algorithm has two modes. Speech mode works with 16 kHz audio and returns MOS-LQO from 1 to 5, with the scale tuned to reach 5.0. Audio mode was designed for music and general audio, and its maximum score is near 4.75. The two scales do not mix: a 4.5 in audio mode is not the same as a 4.5 in speech mode.
In practice
| Mode | Use | Sample rate | Maximum score |
|---|---|---|---|
| Speech | Voice, telephony | 16 kHz | 5.0 |
| Audio | Music, general audio | 48 kHz | about 4.75 |
On a call using the G.711 codec, audio is sampled at 8 kHz. To use speech mode, the reference and the received audio are brought to 16 kHz before the comparison. The reference has to be the same content that should arrive: if the destination plays a greeting, the reference is that greeting, recorded. If the received audio is something else, such as a carrier message, similarity drops and the score follows.
How Monitor uses this
Monitor uses ViSQOL in speech mode, and says so openly so that the way the score is calculated can be checked. On every answered test call, it compares the received audio with the expected recording, the reference recorded when the circuit is set up. Defaults: from 3.8 up the call is healthy; below 3.0 it counts as degraded; between 3.0 and 3.8 it passes with a note. See voice quality with MOS-LQO.
Frequently asked questions
Is ViSQOL free?
Yes. ViSQOL is published by Google Research as open source, in the github.com/google/visqol repository. Anyone can download, build and run the algorithm on their own audio files to check a score. That sets it apart from POLQA, the ITU-T P.863 standard, which is distributed under a commercial license.
Which ViSQOL mode should I use for phone calls?
Speech mode, which works with 16 kHz audio and returns MOS-LQO on a 1-to-5 scale, with a ceiling of 5.0. Audio mode was built for music and general audio at 48 kHz, and its maximum score is near 4.75. For telephone voice, speech mode is the right one.