A program judges speech the way a listening panel would, without seating the panel. It keeps a clean recording, compares that recording with the same words after a codec or a network has altered them, and returns a score meant to predict how people would rate the degradation.
The result is an automated stand-in for a Mean Opinion Score (MOS). People already use this listening scale when they rate speech, and the automated result is mapped onto that same scale.
How the Comparison Runs
- Reference sample. The original speech, before the codec or the network touched it.
- Degraded sample. The same speech after that path.
- Predicted rating. A score on the listening scale people already use, rather than a separate engineering unit.
The Reference Requirement
The comparison cannot run on a degraded file alone. Without the clean original, a quiet passage or a speaker’s accent cannot be separated from damage added on the path. The sender in a codec trial or a network trial still has that original, which is why the comparison fits those tests.
A later method, Perceptual Objective Listening Quality Analysis (POLQA), also predicts a listening score from a reference and a degraded sample. It is used when the audio spans a wider set of conditions than a traditional telephone call.
Human Score and Automated Score
| Question | Human Listening Score | Automated Comparison |
| Who rates it | A panel of listeners | A program |
| What is required | The speech they are asked to rate | The clean original and the degraded sample |
| What comes back | A score on a shared listening scale | A score mapped onto that same scale |
| How it is repeated | A new session with listeners | The same two files, run again |