Skip to main content

Summarization and structured analytics

When Transcription Analytics is enabled for an account, Speechlytics analyzes the completed transcript and adds a summary plus structured conversation fields to the call result.

Summary types​

The account chooses one of two analysis profiles:

ProfileOptimized for
SalesSales performance, objections, business value, buying signals, outcomes, and next steps.
Customer SupportIssue categories, urgency, resolution, follow-up requirements, retention risk, and service quality.

The selected type controls the prompt used to analyze new calls and how agent results are aggregated. Changing it affects future processing; it does not reinterpret previously stored calls.

Speechlytics can also configure a custom analysis prompt for an account whose needs go beyond the two standard profiles.

One-party transcripts​

When only one side of a call is transcribed (mono audio, a silent channel, or an account set to transcribe only the left or right channel), Speechlytics analyzes the call with prompts written for an agent-only transcript. The analysis then focuses on what the agent said and does not invent the customer's side.

Language of the analysis​

Analysis is written in English by default. With native-language analytics, the summary and other written findings are produced in the call's own language, while the classification labels stay in English.

Result fields​

callAnalysis holds the structured analysis in nine sections. In REST responses the keys are the labels below, spaces included (for example "Call Metadata"); webhook and call-event payloads use compact names instead (CallMetadata, ScoringInsights).

SectionFields
Call MetadataAgent, Customer, Call Type
Conversation SummaryMain Topics Discussed (list), Resolution / Next Steps
Customer SignalsPositive Indicators (list), Potential Concerns (list), Buying Intent Level, Urgency Level
Agent PerformanceStrengths (list), Opportunities for Improvement (list)
Sentiment, Topics & EmotionTopics (list), Overall Sentiment, Customer Emotion, Agent Emotion
Business InsightsSales Opportunity, Upsell / Cross-sell Opportunities, Issue Category, First Contact Resolution, Escalation Signals, Customer Retention Concerns, Follow-up Actions Required, Risk Level, Churn or Objection Signals, Next Steps (Critical)
Extracted Keywords & EntitiesKeywords (list), Entities with Names, Companies, Products, Dates, Contact Info
Scoring InsightsRatings out of 5 (see below)
Quick Insight SummaryA one-paragraph takeaway

The object also carries callId and summaryType (0 Sales, 1 CustomerSupport). Which fields are filled depends on the profile: a support call typically leaves sales fields empty, and the reverse.

Scoring insights​

Each rating is a string such as "4/5":

RatingProfile
Customer Satisfaction, Agent Clarity, Professionalism & Tone, Resolution EffectivenessBoth
Sales Opportunity Management, Sentiment & Relationship BuildingSales
Issue Resolution Quality, Empathy & Active ListeningCustomer Support
Overall Call ScoreBoth. Becomes the call's analyticsScore (4/5 = 80).

The classification fields (Overall Sentiment, emotions, Buying Intent Level, Urgency Level, Risk Level, Issue Category, Call Type, Sales Opportunity) use a stable set of English labels, also when native-language analytics is on, so they can be counted across calls.

Completed call responses can contain:

  • callSummary for the plain-language summary;
  • callAnalysis for scores and structured analytics;
  • topics and sentiments for classification results;
  • analyticsScore for the overall call score, converted from a 1–5 rating to 0–100.

Use GET /api/v1/calls/{id}/status for one call or GET /api/v1/analytics for paginated analytics. The exact shape is documented in the API reference.

Cloud and local processing​

Speechlytics deployments can run analysis on a cloud model or on a private, locally hosted model (the Speechlytics Agent). The local option keeps transcripts inside the deployment and is also used for custom-question grading, contextual sensitive-content detection and transcript polishing. The processing location does not change the API contract.

See Analysis profiles for how the profile shapes results.

Processing order​

Summarization runs after transcription and, when enabled, after sensitive-content redaction. This prevents unmasked secrets from being passed into downstream keyword and analytics stages. A call is analyzed once; when a transcript is polished or reverted, analytics are recomputed.

If analytics is disabled, the transcript can still complete successfully without callSummary or callAnalysis.