Scoring Modes

audio_samples_qoe.visqol() supports two modes, selected with the mode keyword argument (default "audio").

Audio Mode

score = visqol(ref, deg)                # default
score = visqol(ref, deg, mode="audio")  # explicit

Property	Value
Operating rate	48 kHz (both signals resampled)
Gammatone bands	32, up to Nyquist
MOS mapping	Support vector regression (libSVM)
Patch selection	All analysis windows
Identical score	~4.73

Use audio mode for music, broadcast content, or any full-bandwidth signal. The SVR model is calibrated on codec and bandwidth-limiting degradations at 48 kHz.

Speech Mode

score = visqol(ref, deg, mode="speech")

Property	Value
Operating rate	Reference’s native rate (degraded resampled)
Gammatone bands	21, capped at 8 kHz
MOS mapping	Exponential NSIM → MOS fit
Patch selection	Voice-activity gated (silent frames excluded)
Identical score	5.0 (exact)

Use speech mode for narrowband or wideband telephony, VoIP, or voice codec evaluation. Upstream recommends 16 kHz input. The exponential fit maps a perfect NSIM of 1.0 to exactly 5.0, so identical signals always score 5.0.

Choosing a Mode

Scenario	Recommended mode
Music quality (streaming, codec, mastering)	"audio"
Speech codec / VoIP / ASR pre-processing	"speech"
Mixed content	"audio"
Narrowband telephony (≤ 8 kHz)	"speech"

When in doubt, use "audio". The SVR model generalises better across content types; speech mode’s voice-activity gating can produce surprising results on non-speech signals with long silences.

Conformance

Both modes are conformance-tested against Google’s C++ reference. Maximum deviation across the test corpus:

Mode	Max Δ MOS	Notes
Audio	0.025	FMA instruction scheduling
Speech (short clips)	0.08	Short-duration fit sensitivity