A process handling 30,000 calls a month with a quality team that audits 400 of them is looking at about 1.3% of its customer conversations. Everything else, the other 98.7%, is known only through averages like handle time and survey scores. Speech and interaction analytics promise to change that ratio by having software “listen” to every call and read every chat and email. The promise is real, but so are the ways it goes wrong, and the useful skill is knowing which questions the technology answers well.
What the software actually does
Under the label “speech analytics” sit a few distinct steps:
- Transcription. Recorded audio is converted to text, usually with separate channels for agent and customer so the system knows who said what.
- Acoustic measurement. Without understanding words at all, the system can measure silence, talk-over (both speaking at once), talk-time ratio, speaking rate and sometimes pitch and energy changes.
- Detection. Rules or trained models look for things in the transcript: a phrase such as “this call may be recorded”, topics like “refund” or “cancel”, signs of dissatisfaction, mentions of competitors.
- Scoring and categorisation. Detections are combined into call categories (why the customer called) and into automated scores against parts of the quality form.
Newer tools built on large language models can summarise calls and judge softer criteria, such as whether the agent explained the next steps clearly. Text channels skip the transcription step, which is one reason chat and email analytics tend to be more accurate than voice.
Where full coverage genuinely beats sampling
Rare but serious events. If a mandatory disclosure is missed on 1 call in 200, a 400-call audit sample will catch about two of them a month. Full coverage finds them all, or at least far more, which matters for compliance-critical processes like collections, insurance and lending.
Contact reasons at scale. Categorising every call by reason tells you that, say, delivery-status calls rose sharply after a courier change. A sample would show it eventually; full coverage shows it in days, with enough volume to break down by region or product. This feeds Voice of Customer reporting for the client.
Agent-level patterns. As covered in discussions of QA sample size, five audits per agent cannot separate good agents from unlucky ones. Automated metrics across all of an agent’s calls, such as silence ratio or hold frequency, give a much steadier picture on the behaviours the system measures reliably.
Smarter human sampling. The best early use is often not replacing auditors but pointing them. Instead of random calls, auditors review the calls the system flagged as long silences, high negative sentiment or possible compliance misses.
Checking whether the machine is right
Every detector makes two kinds of mistake: flagging calls that are fine (false positives) and missing calls that have the problem (false negatives). Vendors may quote an overall accuracy; what matters is how a detector performs on your calls, with your accents, your languages and your scripts. Measuring it is straightforward.
Suppose a “disclosure missed” detector flags 900 of 30,000 calls. The quality team reviews 200 of the flagged calls and confirms 140 genuinely missed the disclosure. The detector’s precision is about 70%: of what it flags, 70% is real. Applied to all 900 flags, that suggests around 630 real misses.
Precision does not tell you what it missed. For that, the team reviews a random 200 of the unflagged calls and finds 4 that also missed the disclosure, 2%. Across the 29,100 unflagged calls that implies roughly 582 missed by the detector. Estimated recall, the share of real problems the detector catches, is then 630 ÷ (630 + 582), or about 52%. The tool is finding only half of the problem while sounding confident about all of it. Without this second check, nobody would know.
Real numbers will differ, and these estimates have their own sampling uncertainty, but the method generalises: validate both directions, repeat periodically, and re-check whenever scripts, languages or the model change.
Where it struggles
- Code-mixed speech. Indian contact centres routinely handle Hinglish, Tanglish and frequent switches between languages within one sentence. Transcription quality on mixed-language and regional-accent audio varies a lot by vendor and has historically lagged single-language English.
- Poor audio. Mobile calls with background noise, low-bitrate recordings and mono recordings that mix agent and customer reduce accuracy.
- Judgement-heavy criteria. Was the resolution correct for this customer’s account? The system often cannot see the CRM or knows the policy only as well as it was configured. Accuracy of resolution, the highest-weighted parameter on many forms, remains hard to automate.
- Sentiment. Sarcasm, politeness norms and a customer who is calm but about to cancel all confuse sentiment scores. Treat them as a sorting signal, not a verdict.
Moving from sampling to full coverage sensibly
A staged approach tends to work better than switching the quality form to automatic scoring on day one. First, use analytics for discovery and contact-reason reporting, where errors are tolerable. Next, automate the objective parameters, such as disclosure statements, greeting, hold notifications and silence, once their precision and recall have been validated. Keep humans on resolution accuracy, complex judgement and coaching. Over time, auditors shift from listening to random calls to reviewing flagged calls, calibrating the machine, and investigating trends.
Two non-technical issues need attention early. Agents will reasonably worry about being scored by software they cannot question, so a clear dispute route and transparency about what is measured matter. And recordings and transcripts are personal data: processing them, storing transcripts and sending audio to a cloud vendor must fit the client’s contract and applicable data protection law, including India’s Digital Personal Data Protection Act, 2023, and any rules on consent, retention and cross-border transfer that apply.
Used well, analytics changes the question quality teams ask from “how did the calls we happened to pick go?” to “what is happening across all our conversations, and which ones need a human to look closer?”
Written by the RP Strativa Team. Volumes and detector figures are illustrative.
Leave a Reply