Sentiment
How caller sentiment is scored turn by turn, what a sentiment drop is, and how to read the trajectory on a scorecard.
Sentiment tells you how the call felt to the caller, independently of whether the agent followed the script. VoiceLens scores it per turn and turns sharp deteriorations into a fault with a timestamp, so coaching can point at the sentence rather than the call.
How it is scored
Each transcript turn receives a sentiment value between -1 (negative) and +1 (positive) from the transcription provider. VoiceLens then builds a customer-only trajectory: the customer's turn scores are smoothed over a three-turn window so a single sarcastic word does not register as a collapse.
The scorecard shows the raw and smoothed customer trajectory as a line, the agent's average for comparison, and the turns where the trajectory shifted by more than 0.2.
| Threshold | Value | Meaning |
|---|---|---|
| Negative | at or below -0.33 | The caller is unhappy. |
| Positive | at or above +0.33 | The caller is happy. |
| Recovery | at or above -0.10 | The caller has come back to neutral after a dip. |
| Shift | ±0.20 | A change between consecutive smoothed values worth marking on the chart. |
The sentiment drop fault
SENTIMENT_DROP fires on the first turn where the smoothed customer sentiment crosses into negative territory after having been neutral or positive. The evidence records three turns: where the dip started, the lowest point, and (if it happened) the recovery.
Its severity depends on the outcome:
- Warning if the caller never recovered to neutral before the end of the call. This costs 5 points on the health score.
- Info if the caller recovered. The moment is still worth looking at, but the agent turned it around, so it does not cost points.
Because the fault is anchored to a turn, the transcript shows exactly what the agent said in the turn or two before the drop. That context is usually the coaching point.
Aggregates
Per call, the API and the scorecard report the customer average, the agent average and the trajectory. Across calls, the channel and agent views average customer sentiment so you can see whether a new prompt, a new script or a new hire moved it.
Reading it well
- A drop that coincides with
MISSED_CORRECTIONorREPETITIONalmost always has the same cause: the caller felt unheard. - A drop with no other fault nearby is often about content (a price, a policy answer) rather than conduct. The agent may have done nothing wrong.
- Recovered drops are good news for coaching: they show what a recovery sounds like on your calls.
- Compare agents on the share of calls with an unrecovered drop, not on the average score. Averages hide the calls that matter.