Sentiment Analysis for Calls: How It Works and What to Expect

Sentiment analysis for calls is software that turns a recorded or live phone conversation into text, then scores how positive, neutral or negative the customer (and the agent) sounded. It looks at the words used and, in better systems, at tone and pitch too. Businesses use the scores to spot unhappy callers early, coach staff on real conversations and see patterns that averages hide. It works best when your phone system captures clean audio, you tell callers about the recording, and you treat the score as a pointer to a conversation worth reviewing, not a verdict.

TL;DR

• Call sentiment analysis transcribes a call, then scores the language and, in some tools, the tone of voice, per sentence, per speaker turn or for the whole call.
• Accuracy depends heavily on the quality of the transcript and the audio. Treat scores as a way to find calls worth a human listen.
• In the UK you must tell callers that calls are recorded and why, and the ICO expects a Data Protection Impact Assessment for higher-risk monitoring.
• The most useful results come from “managing by exception”: reviewing the flagged calls rather than sampling at random.
• A reliable VoIP line, solid broadband and a well-configured phone system are the foundation, because poor audio produces poor scores.

Table of Contents

• What call sentiment analysis actually is
• How the software reads words and tone
• What businesses learn from it
• A worked example for a small business
• How accurate is call sentiment analysis?
• Can ChatGPT analyse sentiment in call transcripts?
• Recording and analysing calls under UK law
• What to look for in a call sentiment setup
• Keep people in the loop
• Our perspective
• Getting your phone system ready
• How we can help
• FAQ
• Sources

What call sentiment analysis actually is

Picture a manager who could listen to every call your business takes and jot down whether the caller ended the conversation happier or angrier than they began. Nobody has the time. Sentiment analysis does a rough version of that automatically.

The process has three stages:

1. Capture. The call is recorded, or streamed live, from your phone system.
2. Transcription. Speech-to-text software converts the audio into words, usually labelling who said what.
3. Scoring. Natural language processing rates each stretch of the conversation. Some tools also analyse the audio itself for tone, pitch and pace.

The output is typically a score on a scale (one UK example is a scale from -1 to +1, which Sky is reported to use to score each customer interaction and find friction points in customer journeys), a positive, neutral or negative label, or a trend line across days and teams.

It is a feature of wider speech or conversation analytics. Sentiment is one signal alongside others such as call length, silence, talk-over, keywords and whether the agent followed a script.

How the software reads words and tone

There are two broad signals, and understanding both explains most of the strengths and weaknesses.

The words (lexical content)

Once the call is a transcript, the software looks for language that carries positive or negative meaning: “brilliant”, “thanks, that’s sorted”, “this is the third time”, “I’m cancelling”. Good tools also weigh context, so that “not bad” is not read as negative.

The voice (acoustic cues)

Tone, pitch and speed carry emotion that words do not. A calm “fine” and a clipped “fine” are the same word. Microsoft researchers found that combining acoustic cues with the words gives more accurate detection than words alone, particularly for negative emotions. That matters for a business, because negative calls are the ones you most want to catch.

The level of scoring

Analysis can happen at the utterance, the speaker turn or the whole-call level, and the individual scores are then aggregated into an overall picture of the customer’s experience. The distinction is practical:

A whole-call score alone can mislead. A call that starts furious and ends warm may average out as neutral, which hides both the problem and the good recovery.

What businesses learn from it

The value is not the score itself but what you do with it. Based on the research and guidance we have read, the uses fall into four groups.

Finding unhappy callers quickly

Businesses use sentiment analysis to identify frustrated customers, flag calls for supervisor intervention in real time and target coaching. For a small company this might be as simple as a daily list of the ten most negative calls, reviewed by the owner over a coffee.

Coaching from real conversations

The Forum, a UK and Ireland customer-contact body, describes how speech analytics directs managers to relevant call snippets so they can shift from managing by average metrics to managing by exception. Instead of listening to five random calls a month, you listen to the five that the data says matter.

Spotting trends

Jet2 uses large language models to analyse intent and sentiment across thousands of call transcripts, which lets it identify trends such as common queries that could be handled through self-service. The same logic works at small scale. If a third of your negative calls begin with “I’ve been trying to reach you”, the problem is call answering, not the agents.

Improving the process behind the call

Negative sentiment often points at something upstream: a delivery delay, a confusing invoice, an engineer who did not turn up. The best use of the data is to fix those causes, which is where moving from a score to a cause becomes the real work.

The Forum also makes a point worth remembering: the power of speech analytics lies in the analyst’s ability to build the right queries, not only in the out-of-the-box features. The tool gives you the data. Someone still has to decide what to look for.

A worked example for a small business

Imagine a 12-person Essex firm that takes around 60 inbound calls a day on a hosted phone system. Here is a realistic way to use sentiment scores without a contact-centre budget.

1. Decide the question. For example: “Which calls end with an unhappy customer, and why?” A clear problem definition is exactly what UK contact-centre leaders recommend before introducing AI, together with rigorous testing.
2. Record and tell callers. Add a short announcement at the start of the call and update your privacy notice.
3. Review the bottom 5%. Each week, listen to the lowest-scoring calls and tag the cause: wait time, missed callback, billing, product fault.
4. Fix one thing a month. If “missed callback” appears most, change the process: call-back rules, a rota, or an automated receptionist that captures the caller’s details.
5. Re-check. After a month, see whether that category shrinks.

Notice what is missing: no one is ranked or disciplined on a score alone. That is deliberate, and the next section explains why.

How accurate is call sentiment analysis?

The honest answer is that it is useful but imperfect, and the accuracy you get depends on your setup. If you are reading a vendor’s headline percentage, ask what audio and what language it was measured on.

Speech is harder than text. Academic work shows that sentiment techniques perform significantly worse on spontaneous, long-form spoken conversations in contact centres than on written text, and that contact-centre conversations differ from other speech, so the tools need adapting to the domain.

The transcript is the foundation. Accuracy depends heavily on transcript accuracy. Even a respectable word error rate can miss the specific polarity-bearing words, such as “not” or “cancel”, that define sentiment. One wrongly transcribed word can flip a sentence.

UK speech recognition has teething issues. The Contact Centre Management Association (CCMA) reported in 2025 that 55% of UK contact centres using automated speech recognition see negative customer sentiment due to inaccuracies. Regional accents, background noise and crackly lines all make matters worse.

Vendors can oversell. The CCMA also notes that sentiment recognition faces significant accuracy challenges and that capabilities are sometimes oversold.

What this means in practice

• Use scores to prioritise which calls a person reviews, not to judge a person.
• Start with a small trial on your own calls, and check a sample of scored calls by ear.
• Invest in audio quality first. A stable connection, a modern handset or headset and properly configured VoIP all raise the quality of what the software hears.
• Choose a tool that can handle your callers’ accents and your industry’s vocabulary.

Can ChatGPT analyse sentiment in call transcripts?

A general-purpose AI chatbot can read a transcript you paste in and describe the mood, and for a handful of calls that can be a handy experiment. The limits are practical:

• It reads text, not voice. Unless the audio has first been transcribed, it has nothing to analyse, and it does not hear tone, pitch or pauses. Given the Microsoft finding above, that means weaker detection of negative emotion.
• It is only as good as the transcript. Every transcription error carries through.
• It is manual. Copying transcripts one by one does not scale to dozens of calls a day, and there is no live alert for a supervisor.
• Data protection applies. Call transcripts contain personal data. Pasting them into a public tool without a proper agreement, and without telling callers, creates a UK GDPR problem. Check the terms of any service with your data protection lead first.
• Consistency. The same transcript can get slightly different wording or ratings on different runs, which makes trend reporting harder.

For a regular, auditable process, a tool built into your telephony or analytics platform, with defined scoring and logs, is the sounder route.

Recording and analysing calls under UK law

Sentiment analysis starts with recording, and recording is where UK businesses need to be careful. This is general guidance, not legal advice. Your data protection lead, or a specialist solicitor, should confirm what applies to you.

Be transparent. The ICO says organisations must tell people that calls are being recorded and explain why. An opening message such as “calls may be recorded for quality and training purposes” is the familiar form, and it needs to match what you actually do. If you also run automated scoring, say so in your privacy notice.

Pick a lawful basis. Many organisations rely on legitimate interests for recording for training and quality control, but they must be able to satisfy the purpose, necessity and balancing tests. Write the assessment down.

Do a Data Protection Impact Assessment (DPIA). For monitoring staff, the ICO’s guidance points to a DPIA, with monitoring that is proportionate and targeted, and with workers’ reasonable privacy expectations managed. If agents are being scored, they are data subjects too, so tell them.

Watch emotion analysis. Using AI to analyse emotion is coming under increased scrutiny, particularly as the EU AI Act outlines. If you serve customers in the EU, take advice on how that regulation affects you.

Keep the data tight. Set a retention period, limit who can listen, and protect payment details if cards are taken by phone.

We cover the practical side of this in more detail in our guide to call recording compliance. If you also track which marketing brings in your calls, this website call tracking guide is worth reading alongside it, because tracking and recording both involve personal data and similar transparency duties.

What to look for in a call sentiment setup

Whatever you choose, judge it against the same checklist.

CCMA guidance adds that any voice automation should reflect your brand tone, escalation policies and customer expectations. A score is only useful if you know what you will do when it is low.

Keep people in the loop

The same CCMA series observes that as contact centres mature, they position human accessibility as a differentiator instead of treating every transaction as a cost to be automated away. For a small business, that is encouraging. Your edge is that a real person who knows the customer can pick up the phone. Sentiment analysis should help that person find the right call at the right time, not replace them.

Our perspective

At Essex Telephone Systems we spend our days on the part of this topic that comes before any analytics: the phone line itself. Most disappointing call data we see traces back to poor connectivity, patchy wifi, or phone systems that were never set up to suit how a business actually answers calls.

Our view is simple. Get the foundations right and everything on top works better. A stable business broadband or leased line, a hosted VoIP system configured around your call flow, and a wifi network designed for the building all give any later analysis cleaner audio to work with. Sentiment scoring cannot fix a call that kept dropping.

We also think small businesses should be wary of buying analytics before they have solved the basics: are calls answered promptly, is there somewhere for out-of-hours callers to go, and does someone follow up? An AI-powered phone system for automated reception and call handling can take care of the first of these, so that your team’s time goes on the calls that need a person.

Because we work without long-term contracts and with UK-based support, you can start with the part that hurts most and add to it as you learn from your own calls.

Paul

Getting your phone system ready

If you want to try sentiment analysis, whether through your telephony provider or a separate tool, prepare in this order:

1. Check the line. Dropouts, echo and delay show up in transcripts as errors. Review your broadband or leased line and consider a 4G/5G backup router so that calls survive an outage.
2. Check the wifi. If staff use softphones or handsets on wifi, weak coverage hurts call quality. A wifi survey shows where the problem spots are.
3. Check the system. A hosted VoIP platform makes recording, call routing and exports far simpler than an ageing on-site box.
4. Write the policy. Announcement, lawful basis, DPIA, retention, access.
5. Pilot on a small group. Compare the software’s scores with a person’s judgement on 20 or 30 calls before relying on it.

How we can help

We supply and install business telecoms for companies in Essex, London and the South East: VoIP and SIP hosted telephony, AI-powered business phone systems for automated reception and call handling, leased lines, wifi survey and design and network cabling. Our business broadband starts from £24 per month for Basic, £38 per month for Standard and £49 per month for Fibre, and 4G and 5G routers start from £20 per month, all without long-term contracts.

Tell us how your calls are handled today and we will suggest where a better line, system or setup would make the biggest difference. You can start on our main site.

FAQ

What is a good accuracy for sentiment analysis on calls?

There is no single trustworthy figure, because accuracy depends on the audio, the accents, the language and how the sentiment was labelled in testing. Spoken conversations are harder to score than written text, and transcript errors flow straight into the scores. Ask any provider to show results on recordings similar to yours, and verify a sample by ear.

Is it legal to analyse customer calls for sentiment in the UK?

Yes, provided you comply with UK GDPR. That means telling callers the call is recorded and why, having a lawful basis such as legitimate interests, and carrying out a DPIA where monitoring is higher-risk. Because rules can change and your situation is specific, check your approach with your data protection lead or a solicitor.

Can a small business use call sentiment analysis?

Yes, and you do not need a contact centre to benefit. A hosted phone system that records calls, plus a simple weekly review of the lowest-scoring conversations, is enough to start. The aim is to find the calls that need a person’s attention, then fix the cause.

Does sentiment analysis replace listening to calls?

No. It narrows down which calls to listen to. Scores can be wrong, especially with poor audio, so a person should always review the calls before drawing conclusions about a customer or a member of staff.

Sources

• Call Center Sentiment Analysis: How It Works on Voice
• Acoustic and Lexical Sentiment Analysis for Customer Service Calls
• Sky customer story: interaction sentiment scoring
• How to gain measurable benefit from Speech Analytics
• Jet2 case study on customer sentiment across call transcripts
• The Hidden Cost of Over-Automation (CCMA)
• Nine Ways Contact Centres Are Currently Using AI (CCMA)
• The Best of Both Worlds: Sentiment Analysis in contact centres
• ICO: monitoring workers, specific considerations
• ICO: what is the legitimate interests basis

Recommended

• AIVoice
• Voip