How sentiment analysis works on calls
After a call, a model reads the transcript and judges the tone, turn by turn and for the call as a whole. Some systems also use the audio: pace, volume and pauses.
On Indian calls, the words alone can mislead:
- "Theek hai, theek hai" can mean agreement, or "please stop calling".
- "Bas karo yaar" is clearly negative, though no word in it is rude.
- Politeness can hide a no: "Ji, main dekh leta hoon" may mean the caller isn't interested.
Take two calls that both end with "callback requested". One caller was relaxed and simply busy; the other was irritated after a third reminder. Sentiment tells them apart, so the second gets a gentler approach, or a person, next time.
Why it matters
A disposition says what happened; sentiment says how it felt. Tracked by agent and by campaign, it shows when a script change has made calls tenser, which campaigns are wearing people out and which calls a supervisor should listen to first.
How Kallix uses it
Every Kallix call is tagged with a sentiment alongside its disposition, transcript and captured fields. Analytics shows sentiment by agent and by campaign, so you can spot a problem early and check it against the recordings.