Skip to content

Transcription & Call Intelligence

Aura AI turns the spoken audio your phone system captures into text and intelligence: transcripts (with who-said-what speaker labels), AI summaries, sentiment, topics & keywords, and semantic search across everything that was said. This page is the map — what gets transcribed, what each analysis layer adds, where each result shows up in Switchboard, and how your plan gates it.

There are two pipelines, and they behave differently:

SourceWhen it runsSetup neededPlan
VoicemailAutomatically, on every deposited messageNoneEvery plan — free
Recorded callsAutomatically on ingest when Auto-process is on, or on demand per callCall recording enabled on the extension + auto-process or a manual requestTranscription on every Nova plan; deeper analyses by tier (see Plan gating)

Very short audio is skipped on purpose — a few seconds of ring, beep, or silence isn’t worth a transcript. Skipped calls can still be transcribed manually from the call detail (see On-demand transcription).

Every voicemail left on a Nova mailbox is transcribed automatically, on every plan, with nothing to configure. The inbox at Nova PBX → Analytics → Voicemail shows the AI results inline — a sentiment badge and keyword chips per row, and a triage line that counts unread and negative messages so unhappy callers surface first.

Open any message to see the full result:

Voicemail detail with transcript, AI summary, sentiment, and topics & keywords

  • Transcript — the message as text, on every plan. Triage your inbox without pressing play.
  • Summary — a one-line “why they called”, when your plan includes summarization.
  • Sentiment — Positive / Neutral / Negative, when your plan includes sentiment analysis.
  • Topics & keywords — chips for fast scanning, when your plan includes NLP processing.

Voicemail is a single speaker, so there is no diarization step — the plain transcript is the whole feature. Messages shorter than a couple of seconds are skipped as hang-ups.

See Voicemail for the full inbox workflow (filters, mark read/unread, email notification, retrieval by phone).

For two-sided conversations, Aura transcribes recorded calls. Recording capture itself is configured per extension; whether recorded calls are processed automatically is the Auto-process recordings with Aura AI switch on Recording Settings:

Recording Settings — Aura auto-process pipeline and the speaker diarization toggle

With auto-process on, every recorded call flows through the pipeline shortly after hangup: transcription → (optionally) speaker diarization → the analyses your plan includes. With it off, nothing is processed until someone requests it per call.

Open any call from Nova PBX → Analytics → Call History, then switch to the Transcript tab:

Diarized call transcript — speaker-labeled turns with timestamps

  • Each speaker turn is labeled (Speaker 1, Speaker 2, …) and color-accented, so a two-party call reads like a conversation instead of a wall of text.
  • Every segment carries a timestamp. Click any line to jump the audio player to that moment — the fastest way to verify what was actually said.
  • Without diarization (lower tiers, or the toggle off), the same tab shows a plain timestamped transcript with no speaker labels.

Speaker diarization — who said what, where

Section titled “Speaker diarization — who said what, where”

Diarization adds who-spoke-when labels to call transcripts. It answers the question a flat transcript can’t: which side of the call said that?

How it works. Diarization separates the recording by voice, labelling each distinct voice as a speaker and stamping every turn with its start time. It runs as queued post-processing on recorded calls — there is no real-time impact on live calls.

What it can do:

  • Label each turn in a two-party call so agent and caller are visually distinct.
  • Handle more than two voices (three-way calls, transfers) — each distinct voice gets its own label.
  • Combine with click-to-seek: jump straight to the moment a specific person said something.

What it can’t do:

  • It doesn’t know names. Labels are positional (“Speaker 1” is whoever spoke first), not identities. It does not match voices to your extension directory.
  • Crosstalk is hard. When both parties talk over each other, segments may merge, swap, or drop words.
  • Hold music and IVR prompts can occasionally register as a “speaker”.

Enabling it. Diarization is a Nova Ultra entitlement with its own toggle on Recording Settings (the server refuses to enable it on lower tiers). Once on, it applies to newly processed recordings.

Per-call insights: summary, sentiment, topics

Section titled “Per-call insights: summary, sentiment, topics”

The Insights tab on the same call detail shows what Aura’s analysis layer extracted from the transcript:

Call Insights tab — AI summary, sentiment label, topics & keywords

  • Summary — a few sentences covering why they called and how it ended. Read it instead of the recording.
  • Sentiment — one of three classes: Positive, Neutral, or Negative, judged across the whole call. Sentiment also appears as a badge in the Call History table’s AI column, so you can spot unhappy calls without opening anything.
  • Topics & keywords — short tags (e.g. billing, invoice 4471) for scanning and cross-call patterns.

Analyses your plan doesn’t include render as locked rows with an upgrade link — the transcript itself is never hidden behind them.

Calls that weren’t auto-processed (auto-process off, call skipped as too short, or a failed run) show a Transcribe this call button on the Transcript tab. One click queues the same pipeline for that call; the tab shows “Transcribing…” and updates itself when the result lands — typically within a couple of minutes.

Section titled “Semantic search & trends — Aura Insights”

Individual transcripts answer questions about one call. Aura AI → Agents → Insights answers questions about all of them:

Aura Insights — semantic transcript search results and sentiment analytics

  • Transcript search matches meaning, not keywords: searching “billing dispute about an overcharge” finds calls where the caller said “the invoice came in wrong” — no shared words required. Each hit shows the matching snippet, the call’s sentiment, and a match score; click through to the full transcript and recording.
  • Sentiment trend charts daily Positive / Neutral / Negative counts across your processed recordings. Watch for negative spikes after a pricing change, an outage, or a new IVR flow.
  • The page also covers per-voice-agent performance — see Aura Insights for the full page reference.

Search runs over the transcripts of auto-processed recordings — calls that were never processed aren’t in the index.

Aura AI is not a separate subscription — it comes with every paid Nova plan, and your Nova tier decides which layers light up:

CapabilityCoreProUltra
Voicemail transcription✅✅✅
Call (recording) transcription✅✅✅
Speaker diarization——✅
Sentiment analysis——✅
AI summaries——✅
Topics & keywords (NLP)——✅
Semantic transcript search——✅

Everything in this table is entitlement-gated, not metered — there is no per-minute charge for transcribing or analyzing your own calls on an entitled plan. In the product, locked rows and upgrade prompts reflect your live entitlements; manage your plan under Account → Subscriptions.

Transcription is tuned for telephone audio, but it is machine-generated:

  • Clear audio with one speaker per turn transcribes very accurately; background noise, low line quality, strong accents, and crosstalk reduce accuracy.
  • English is the primary supported language.
  • Transcripts are not human-reviewed. For legal or compliance purposes, treat the transcript as an aid and the audio recording as the authoritative record.

Transcripts and AI analyses are derived from — and live with — the underlying audio:

  • Retention follows the recording/voicemail. Deleting a recording or voicemail removes its transcript and analyses; recordings age out per your plan’s retention window (see Recording Settings → Storage and retention).
  • Access is permission-gated. Viewing call transcripts requires call-history access (nova.cdr.view_all); listening to the audio additionally requires the Listen to Recordings permission. Transcript search is hidden from users without call-history access, since snippets expose caller numbers.
  • Consent. If your jurisdiction requires consent to record, that consent covers the transcription and analysis of those recordings. The recording disclosure announcement is configured on Recording Settings — and remember on-demand *9 recordings play no announcement.
SurfaceWhat you’ll findPath
Voicemail inboxPer-message transcript, summary, sentiment, keywordsNova PBX → Analytics → Voicemail
Call History detailDiarized transcript (click-to-seek), per-call insights, on-demand transcribeNova PBX → Analytics → Call History
Aura InsightsSemantic transcript search, sentiment trends, agent performanceAura AI → Agents → Insights
Recording SettingsAuto-process switch, diarization toggle, disclosure audioNova PBX → System → Recording Settings