The first objective speech quality measure failed on mobile-network timing variations

Evidence

The PSQM was primarily focused on identifying the quality impact of speech codecs. Unfortunately, the method did not turn out to be successful, and the recommendation has since been withdrawn. The PSQM was unable to estimate correctly the quality impairment due to filtering, variable delay, and distortion common in mobile communication. For example, the transmission delay may vary in mobile transmission, and the variations can be fatal in VoIP applications. Subtracting timevarying feature vectors, as shown in Figure 17.5, would estimate a large error in the case where the test case had random variations in the lengths of syllables in speech. Such small delay variations do not cause much deterioration of perceived quality, implying that the model would underestimate the speech quality in this case. — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 2445

Categorization

• XREF: Connects to the broader story of why objective metrics fail in real-world conditions, echoing the FAHQMT debate and listening-test methodology threads.
• QUESTION: Why did subtracting time-varying feature vectors cause the model to misconstrue syllable-length jitter? Worth chasing the PSQM algorithm details.