Glossary

Acoustic measures — properties derived from the audio rather than the transcript: silence, talk-over, talk ratio, speech rate, volume. More reliable than transcript-derived measures because they do not inherit recognition error.

Agreement (inter-rater) — the extent to which evaluators scoring the same interaction produce the same result. The measure that determines whether human scores mean anything.

Calibration — the process of measuring and reducing evaluator disagreement, usually through periodic sessions on shared interactions.

Categorisation — assigning interactions to topics, by rules, by a trained classifier, or by a language model. Its precision and recall should be measured on your own data.

Diarisation — determining who spoke when, from a single mixed channel. Substantially less reliable than stereo capture.

Emotion inference — claiming to detect a person's emotional state from voice. Scientifically contested and increasingly regulated in employment contexts.

First contact resolution — a family of differently-defined measures, most of which do not measure resolution. Repeat contact rate with a stated window and matching rule is the usable version.

Full coverage — analysing every interaction rather than a sample. The main reason to deploy speech analytics, and it changes the volume problem rather than removing the need for judgement.

Keyword spotting — searching audio for sound patterns without producing a transcript. Fast, cheap, and limited to terms configured in advance.

PCI scope — the condition of holding payment card data, which brings recordings and derived objects under payment card security requirements.

Precision and recall — of a category or an automated item: the proportion of matches that are correct, and the proportion of true cases that were caught. Both should be measured and reported alongside any count.

Redaction — removing sensitive content from a recording after capture, as distinct from pause-and-resume, which prevents capture.

Scorecard — the instrument defining what is evaluated. Should separate compliance items, scored pass or fail, from behavioural items.

Sentiment analysis — classifying language as positive, negative or neutral. Measures language, not feeling, and inherits transcription error.

Stereo capture — recording agent and customer on separate channels. The single highest-return technical decision in a speech analytics deployment.

Voiceprint — a biometric template derived from voice, distinct from the recording and regulated more strictly in several jurisdictions.

Word error rate — the proportion of words wrong in a machine transcript relative to a human reference. Varies by accent, audio quality and vocabulary, and should be measured on your own calls rather than taken from a vendor benchmark.

Terms used loosely elsewhere

"Accuracy" in vendor material usually means agreement with the vendor's own reference set, on the vendor's own audio. These notes avoid the word without a stated test set.

"Real time" means a short latency, not live.

"AI-powered scoring" may mean rules, a classifier, or a language model, with very different explainability. Ask which.

"100% quality monitoring" means every interaction is processed by automated checks, not that every interaction is evaluated for quality. The distinction matters and it is frequently blurred.

"Customer sentiment" is a classification of words, presented as a statement about a person.

How these notes use the contested terms

Where the industry is inconsistent, these documents pick one meaning and state it.

"Quality" here means the interaction achieving its purpose for the customer and meeting the operation's obligations. It does not mean conformity to a scorecard, which is what a score measures.

"Coverage" means the proportion of interactions processed by automated analysis, not the proportion evaluated for quality by a person.

"Score" always refers to a human or automated evaluation against a scorecard, never to a survey result.

"Compliance" means the regulated or mandatory requirements, scored pass or fail, kept separate from behavioural assessment throughout.

"Analytics" means the automated layer; "evaluation" means the human one. Where a note says one, it does not mean the other.

"Agent-level" findings are about a person; "process-level" findings are about the operation. A substantial share of what QA discovers is the second, and the distinction is made explicitly in every note where it matters.