Sentiment Analysis Software
How sentiment analysis software works
Sentiment analysis software turns a recorded or live conversation into a score in four steps. Speech-to-text transcription converts the audio into text. Speaker separation, also called diarization, assigns each part of the transcript to the person who said it, so the customer’s words and the agent’s words are scored apart. The transcript is then split into segments, and each segment gets a polarity: positive, negative or neutral. Finally the segment scores are aggregated into an overall score for the call.
What the software reads decides what it can catch. Text-based sentiment analyses the words: “this is the third time I’m calling” is negative whatever the tone. Acoustic sentiment analyses how the words were said, through features such as pitch, volume, speaking rate and silence. Text misses a flat “fine” said through gritted teeth. Acoustic analysis catches it, but it processes voice characteristics, which makes it a different kind of data with different rules attached.
Sentiment trajectory is the change in sentiment from the opening segments of a call to the closing ones. An overall score averages the two, so a call that starts angry and ends satisfied can come out neutral. The trajectory is the measure that shows whether the conversation fixed something.
Sentiment analysis vs. CSAT: what is the difference?
CSAT (customer satisfaction score) is declared. The customer answers a survey after the interaction and gives a score. It is explicit and it is the customer’s own word, but it exists only for the customers who chose to answer.
Sentiment is inferred. The software reads it from the conversation itself, for every call that was analysed, without asking anyone anything. It covers the whole volume, but it is a probabilistic reading: a model’s interpretation, not a statement.
The practical consequence is what a divergence between the two tells you. A queue where CSAT holds steady and sentiment falls has a problem in the calls whose customers did not answer the survey, which CSAT alone would never show. A queue where sentiment is negative and CSAT is high usually has angry callers and agents who fix the issue, which is the trajectory at work. Both sit among the contact center metrics worth tracking together, because neither replaces the other.
What sentiment analysis software is used for
Sentiment analysis use cases split into two groups, and the dividing question is when the score is read: during the call, or after it.
During the call. This needs streaming transcription and someone ready to act on the score.
- Alerting a supervisor when a call turns negative, while there is still time to join
- Prompting the agent with guidance when the customer’s tone changes
After the call, on recordings. This needs a review process, not a live one.
- Finding the negative calls for quality review, instead of sampling recordings at random
- Coaching with the exact segment where the tone changed, not with the whole call
- Tracking sentiment by queue, topic or keyword over weeks, to see where it is moving
- Checking the effect of a change, such as a new IVR menu, on the calls that follow it
- Comparing calls handed over by an AI voice agent with calls that reached a person directly
Most contact centers start with the second group, because it works on recordings they already have. It is one of the post-call steps of contact center automation, and its value depends on someone reading the results every week.
Where Imagicle fits
Imagicle Voice Analytics adds sentiment analysis to the calls recorded by Imagicle Call Recording, as part of Compliance & AI Quality Recording. It transcribes each recording with automatic language detection and scores sentiment as positive, negative or neutral, for the whole call and for each segment. Calls can be filtered by sentiment and tagged by keyword, so the negative ones reach quality review first. Over 80,000 users record their calls with Imagicle.