Speech that has passed through a network can be scored without seating a new panel of listeners for every sample. A later objective model predicts that listening score by holding a clean reference next to the signal the listener actually heard.
The result is mapped onto a human listening scale, reported as a Mean Opinion Score (MOS). The same comparison was already the job of the earlier speech-comparison model, Perceptual Evaluation of Speech Quality (PESQ), which was built for a narrower band of speech. The later model keeps that reference comparison and also scores wider-bandwidth speech.
What the Model Compares
- Clean reference. The undistorted speech sample held back so the model has something to compare.
- Heard signal. The same speech after it has passed through the network, as the listener’s ear received it.
- Predicted score. An objective estimate of the listening judgment for that pair, produced without assembling a new human panel for the sample.
Earlier Model and Later Model
Both score speech that has passed through a network, both compare a reference with what the listener heard, and both map that comparison onto the human listening scale. The later model was built to cover wider-bandwidth speech than the earlier model was built for.
| Comparison | Earlier speech-comparison model | Later listening model |
| Reference check | Compares a reference with what the listener heard | Compares a reference with what the listener heard |
| Speech that is scored | Speech that has passed through a network | Speech that has passed through a network |
| Bandwidth | The narrower band the earlier model was built for | Wider-bandwidth speech than that earlier model was built for |
| Reported result | Mapped onto a human listening scale | Mapped onto a human listening scale |