Speech Analytics & Quality Management

Speech Analytics and Quality Management (QM) extend QA Analytics for voice-heavy contact centers. Where QA scores transcripts on a rubric, Speech Analytics adds the audio layer: dual-channel recordings, sentiment over time, anomaly detection, and a dedicated quality department interface.

What’s in Speech Analytics

Capability	What it gets you
Dual-channel recording	Customer and agent audio recorded on separate tracks. Listen to either side in isolation.
Sentiment over time	Per-second sentiment graph across the call; spike/drop detection.
Verbal anomaly detection	Surfaces calls with raised volume, talk-over, long silences, profanity, or scripted compliance phrase missed.
Editable transcripts	Reviewers can correct STT errors; corrections feed back into model fine-tuning.
Sync transcript ↔ audio	Click a transcript line to jump to that audio moment.
Real-time threshold alerts	Fire a Slack/PagerDuty alert when sentiment drops below a threshold or compliance phrase missed.

Quality Department interface

The Quality Department UI is purpose-built for QA reviewers:

Call list — filter by date, channel, agent, team, sentiment, language, intent, custom KPI.
Listen panel — synced waveform + transcript + agent-customer turn separation.
Assessment cards — fill out a rubric while listening; auto-save per criterion.
Side-by-side calibration — compare your scores with the AI grader’s; flag disagreements.
Coaching action — directly assign a remedial training scenario from the call.

This is a separate workspace from the inbox so reviewers can focus on QA without distraction. Reviewers see only conversations they’re assigned or sampled.

Dual-channel recording

For customer/agent isolation:

Channel	How it captures dual-channel
Genesys	SIPREC media stream is dual-channel by default.
Twilio	Set `recordingChannels=dual` in the start verb.
RingCentral	Native dual-channel recording.
Zoom Phone	Mono only — single mixed channel; AI separation can split agent vs customer with ~95% accuracy.

Once recorded, the per-side audio is exposed in playback as separate volume sliders + waveforms. Click “agent only” or “customer only” to isolate.

Verbal anomaly detection

The anomaly detector runs over each call and flags:

Anomaly	Trigger
Raised voice (customer)	Sustained volume above customer baseline.
Talk-over	Both sides speaking simultaneously > 3s.
Long silence	> 10s without speech.
Required script missed	Compliance phrase (“This call may be recorded…”) not detected within first 30s.
Profanity	Configurable list per locale; flagged but not auto-redacted (preserves audit).
Sentiment cliff	Sentiment drops > 2 standard deviations within 30s.

Anomalies appear as flags in the Quality Department UI and can fire real-time alerts.

Real-time threshold alerts

Configure under Settings → Notifications → Real-time alerts:


Trigger: live conversation
Condition: customer_sentiment < -0.5 for > 30 seconds
Action: post to #cs-escalations with conversation link + audio clip

Common patterns:

Drop in CSAT → notify supervisor mid-call so they can warm-transfer in.
Spike in escalation requests → notify operations.
Agent missed the consent prompt → coach immediately.

Alerts fire to Slack, Teams, PagerDuty, or webhook. See Notifications.

Editable transcripts

The STT layer is good but not perfect. Reviewers can:

Click a transcript line.
Edit the words.
Save.

The edit is logged in the trace, and the per-tenant STT model receives the correction as a training pair. After enough corrections, you can request a tenant-specific STT fine-tune (subject to data volume and contract).

Edits change the transcript, not the audio. The audio is immutable for compliance. The transcript is the human-curated companion.

Bedrock and BYO model support

For customers who require AWS-native inference:

Surface	Bedrock support
Speech Analytics scoring	Available — Claude on Bedrock.
Agent Assist	Available — Claude on Bedrock or Llama via Bedrock.
Realtime voice	Not via Bedrock; use OpenAI / Gemini / Cascaded.

Configure under Settings → AI Infrastructure → Custom model.

Reporting

Speech-Analytics-specific dashboards complement QA reports:

Sentiment distribution per agent and team.
Talk-listen ratio per agent (do they let the customer speak?).
Average longest silence per call.
Compliance phrase coverage rate.
Top verbal anomalies by frequency.

Open in Omniflow

If you want to…	Go to
Score conversations on a rubric	QA Analytics
Send real-time alerts	Notifications
Fine-tune the agent on coaching insights	AI Insights
Read the runtime architecture	Voice Runtime