Manual quality review samples a tiny fraction of contacts, often one or two per agent per week. Speech analytics analyses the whole population. That difference in coverage is the entire value proposition, and everything else about the technology follows from it.
It is also a category where buyers are sold a great deal. Dashboards look impressive in a demonstration and are quietly ignored six months later. This article explains what the technology actually does, what it detects that sampling cannot, where it disappoints, the consent questions it raises, and how to decide whether it belongs in your programme.
What speech analytics is
Speech analytics is software that turns recorded or live calls into text, then searches, categorises and scores that text at scale. It works on every call rather than a sample. The output is a set of findings: which calls contained a given phrase, which agents skipped a required statement, which topics are rising week on week, and which conversations showed the language that tends to precede a complaint.
It sits alongside, not instead of, the quality programme described in our call centre analytics service. Manual review still decides whether an individual agent handled a call well. Analytics decides where to look, and finds patterns that no reviewer could see one call at a time.
There are two modes. Post-call analytics processes recordings after the fact and is where most programmes start, because it is simpler to set up and the findings arrive in a daily or weekly report. Real-time analytics listens as the call happens and prompts the agent or alerts a supervisor while the customer is still on the line. Real-time is more demanding to configure and to act on, and it only pays off where an intervention during the call changes the outcome, such as a missed disclosure on a sales call or a customer about to cancel.
How it works in practice
The process has four stages. Transcription converts audio into text, either after the call or as it happens. Categorisation applies rules or models to tag each call with topics, outcomes and events: a cancellation request, a mention of a competitor, an apology, a disclosure read in full. Search lets a supervisor find every call matching a phrase or a combination of tags. Scoring applies a rubric across all calls to produce a number per agent, per team or per contact type.
Some platforms add acoustic measures such as talk-over, silence and pace, and some produce a sentiment estimate from word choice and tone. Accuracy varies with audio quality, accent, vocabulary and configuration, and any figure a vendor quotes should be tested on your own calls rather than taken from a brochure. Treat every output as a signal to investigate, not a fact.
What it detects that sampling cannot
Compliance gaps. Required disclosures that were skipped and prohibited statements that were made, across every call rather than the handful someone happened to listen to. For a programme with mandatory scripts, this is the first and clearest use: a list of the calls where the script was not followed, ready for review the same day.
Emerging problems. A sudden rise in a phrase, a product name, an error message or a competitor surfaces a new issue in its first week rather than in next month's report. The contact centre usually hears about a problem before anyone else in the business does, and analytics is how that early warning becomes visible.
Escalation predictors. The language patterns that reliably precede a complaint, a chargeback or a cancellation, which lets a supervisor intervene on the call or shortly after it rather than reading about it in a survey. Silence and repeated hold are often as telling as any word.
Process friction. Long hold times on a specific question, agents reading from a page that customers cannot follow, or the same clarification asked on every call all point at a fix in the process or the documentation rather than in the agent.
Where it disappoints
It tells you what was said, not why. A rise in cancellation language tells you customers are leaving. It does not tell you whether that is price, a product defect or a competitor's campaign. Someone still has to listen and think.
Sentiment scoring is approximate. Sarcasm, regional expression and cultural difference all confuse it, and a score presented with two decimal places is no more precise for having them. Use sentiment to rank calls for a human to review, never to judge an agent on its own.
It produces findings, not fixes. Without a named owner for each category of finding, and a review rhythm, it becomes another dashboard nobody acts on. The subscription continues and the behaviour does not change.
Analytics that nobody is accountable for acting on is a subscription, not a capability.
Recording, consent and data handling
Speech analytics depends on recording calls, and the rules on recording and consent vary by jurisdiction. Some places require every party to consent, others require one, and the position for calls that cross borders is not always obvious. Announcements, the wording used, and whether a customer can opt out all need confirming with your counsel for each country and state you serve.
The transcripts themselves are data. They may contain payment card details, health information or other sensitive content depending on the work. Where cardholder data is spoken on a call, PCI DSS governs how it is handled, and pause-and-resume recording or redaction is the usual approach; our article on PCI compliance for phone payments covers the options. Where protected health information is involved, a vendor processing recordings on your behalf is a business associate and needs a business associate agreement in place, with access limited to the minimum necessary. Ask where transcripts are stored, who can search them, how long they are kept and how they are destroyed.
When it is worth buying
Speech analytics earns its cost at volume, in regulated environments, or where the cost of a missed compliance failure is high. A programme handling a large volume of calls each month with disclosure requirements will usually justify it. A programme where a single mis-statement creates legal exposure, such as collections or financial services, justifies it on risk alone.
Below a modest volume, a disciplined manual sampling process usually gives more useful insight per unit of money and management time than a platform will. The threshold is not a fixed number of calls. It is the point where the questions you need answered cannot be answered by listening to a sample, and where someone has the time to act on what the analytics finds.
Before deciding, write down the three findings you would act on if you had them tomorrow, and who would act. If the list is hard to write, the platform will sit unused. If the list writes itself and the actions are clear, the case is usually made. Also decide whether the platform is bought by you or provided inside an outsourced programme, because the second option puts the configuration and the review rhythm on the provider rather than on your team.
How to run a pilot
Start with two or three specific questions rather than switching everything on. Good first questions are narrow and checkable: whether the required disclosure was read in full on every sales call, which topics drive repeat contacts, and which calls contained cancellation language in the past week. Configure the categories for those questions, run them on your own recordings, and check a sample of the results by ear to see how often the tag is right.
Assign an owner for each finding type before the pilot starts. Agree what action follows a finding: coaching, a script change, a documentation fix, a product escalation. Review the results monthly and decide whether the findings changed anything. If they did not, either the questions were wrong or nobody had time to act, and both need fixing before the scope widens.
- Two or three narrow questions, not the full feature list
- Tested on your own recordings, with results checked by ear
- A named owner and an agreed action for each finding type
- A monthly review that asks what changed as a result
- Widen scope only after the first questions are producing action
Questions to ask a provider
Ask how transcription handles your accents, your product vocabulary and your audio quality, and ask to test it on your calls. Ask how categories are built and who maintains them as your scripts change. Ask what is redacted, where data is stored, who can access it, and how retention and deletion work. Ask what the reporting looks like for a supervisor on a Monday morning, not in a sales demonstration. And ask what the provider expects you to do with the findings, because a vendor who has thought about that will describe a review rhythm rather than a dashboard.
Where it fits with the rest of the programme
Analytics is one input to a quality programme, alongside a written scorecard, calibrated reviewers, customer feedback and the operational numbers covered in our guide to the KPIs that matter. It makes manual review more targeted and makes compliance visible across the whole population of calls. It does not replace the person who listens, coaches and fixes the process. See our speech analytics service for how we configure and run it inside a managed programme, with the reporting rhythm agreed up front.