Post-call analytics on a voice stack used to mean "find the recording, hope someone tagged it." The new Voice Intelligence page in Devotel Orbit replaces that habit with a single dashboard that reads your analysed transcripts and folds the whole window into one view: how sentiment is trending, which keywords kept surfacing, which topics your callers actually raised, which intents dominate, and how much your agents talked versus listened. This post walks the shipped page end-to-end — what each panel measures, where the data comes from, and how operators turn the reading into an action.
Why speech analytics belongs on the dashboard, not in a recording archive
A call recording is evidence; a speech-analytics roll-up is direction. The two answer different questions. The moment a supervisor hears "complaints are up," they need the direction — which topics are driving it, which keywords cluster around it, whether the average sentiment per day is actually dipping or one anomalous call is skewing the read. Without an aggregate surface, every one of those answers is a CSV export over per-call fields someone remembers to tag.
Voice Intelligence puts the aggregate in front of the per-call layer:
- Sentiment over time — a day-by-day trendline of average call sentiment, so a dip is visible before it becomes a complaint pattern.
- Top keywords — the words and phrases that recurred most across analysed transcripts in the window, sized by frequency.
- Topic clusters — the subjects your callers actually raised, grouped rather than keyword-matched, with count and average sentiment per cluster.
- Call intents — the auto-detected reason-for-call breakdown (billing question, booking, support issue, and so on), counted against its own denominator.
- Conversational dynamics — the talk-listen ratio, dead-air, and overtalk/interruption averages across the same window, so coaching conversations start from an objective baseline.
Every panel drills down. Clicking a dip day, a keyword, or a topic cluster opens a filtered sample of the calls behind the number, so the dashboard stays honest — the aggregate is always one click from the underlying transcript.
What the page actually renders, panel by panel
The page header owns the window selector — 24 hours, 7 days, 30 days, or 90 days — and that window applies to every panel below. Three headline cards sit under it: total calls analysed, average sentiment across the window, and the count of distinct topics identified. The cards answer the first question a supervisor asks ("is this a busy window or a quiet one?") before they read a single trendline.
The sentiment over time chart plots one point per day — the average sentiment score of the calls analysed that day. A dip is actionable: click the day and a dialog lists the lowest-scoring calls in that sample, so the supervisor reviews the three worst transcripts rather than a day of tape.
The conversational dynamics panel reports four coaching numbers across the window: the talk-listen ratio (how much of the conversation the agent filled versus the customer), the average agent talk share as a percentage, the average dead-air (with the longest single silence for context), and the average interruptions per call. Dead-air and overtalk need timed transcripts, so those two averages cover a smaller call population than the talk-listen ratio — the page states that distinction explicitly in the panel caption rather than letting the numbers be read against the wrong denominator.
The top keywords cloud sizes each recurring term by frequency. Clicking any word filters the call list to transcripts that contain it — the difference between "we keep hearing about cancellations" and "here are the nine calls where cancellation came up."
The topic clusters table groups callers' actual subjects rather than matching a fixed keyword list. Each row carries the topic name, the number of calls in the cluster, and the average sentiment inside it, coloured as positive, negative, or neutral. Clicking a row opens the ten most recent calls about that topic, so an emerging subject gets reviewed while it is still emerging.
The call intents breakdown auto-classifies each completed call into a reason — billing question, booking, technical issue, and so on — and counts them. The count deliberately reconciles against the intent-classified population, not against the top "calls analysed" card, because intent classification and sentiment scoring run on different pipeline stages. The page spells that out in a caption directly above the table so the two totals are never compared against each other by accident.
Turning classification on — categories and custom operators
Two manager sections follow the panels and control what the post-call pipeline computes on every analysed transcript.
Speech-analytics categories are the shared taxonomy a tenant defines once — the business-meaningful call-reason categories the post-call classifier should assign to every call. The classifier short-circuits when the tenant has zero categories, so this section is where classification gets turned on: define a category, and every analysed call from that point is classified against it.
Custom intelligence operators are the tenant-authored attributes the post-call pass computes on every analysed transcript — a yes/no question ("did the agent offer an alternative?"), a category ("which product area was this about?"), or a free-text extraction ("what was the caller's main objection?"). Each active operator drives a billable analysis pass, so admins create, edit, and deactivate them from the same page rather than calling support.
Both sections live in the middle of the page rather than a separate settings area — the operator activating the classifier is usually the same operator reading its output below.
Where to look next
- The Voice Intelligence dashboard — the shipped page this post announces, on the dashboard itself.
- Voice of Customer on a CPaaS — how CSAT, NPS, and CES survey scores land on the same customer record voice analytics reads from.
- Best call recording and voice analytics platforms — the cross-platform comparison if you are still shortlisting.
Frequently asked questions
Where does the sentiment, keyword, and topic data come from?
The page reads the post-call analysis pipeline that processes every recorded call transcript on your tenant — the same pass the speech-analytics categories and custom intelligence operators configure. It is a roll-up of already-analysed transcripts, not an export.
Who can open the Voice Intelligence page?
The page is gated to owner, admin, and developer roles, matching the other voice-operations pages — transcripts are the highest-sensitivity artefact on the platform and access follows that.
What do the conversational-dynamics numbers need to populate?
Timed transcripts. Talk-listen ratio and agent talk share populate on every analysed call; dead-air and overtalk need per-word timing, so those two averages cover the subset of calls whose transcripts carry timing information. The page states that distinction in the panel caption.
Can I drill from a topic or keyword into the underlying calls?
Yes — clicking a dip day, a keyword, or a topic row opens a filtered sample of the calls behind it. Sentiment dips return the lowest-scoring calls; topics and keywords return the most recent calls containing them.
How do I turn intent and topic classification on for my tenant?
In the Speech-analytics categories section on the same page: define at least one category and the post-call classifier starts assigning every analysed call against it. The Custom intelligence operators section on the same page adds tenant-authored attributes beyond the built-in taxonomy.