Speech quality can be predicted algorithmically using signals that mimic human auditory processing.

Evidence

The method is a relatively complex signal processing structure, which includes parts mimicking human time–frequency resolution, and ultimately uses a trained pattern recognition algorithm to produce final quality estimates. It requires three inputs to the system: a clean speech signal, the unprocessed signal, and the processed signal. The clean signal is the speech signal of a real human subject recorded in a free field. The unprocessed and processed signals are measured in realistic noisy conditions created in the laboratory. A dummy head with a mouth is placed in a noisy environment generated in a laboratory, and the device under test is placed near the dummy head, just as it would be in practice. The noisy environment is generated using an equalized loudspeaker set-up with four loudspeakers and one subwoofer around the head. — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 2455

Categorization

• DEFINE: Explains how 3QUEST predicts subjective speech quality using a pipeline mimicking human time-frequency resolution plus a trained pattern-recognition model.