Genesys Cloud speech and text analytics: the operating handbook

2.1 Licences and transcription minutes

STA sits with the workforce engagement family. On Genesys's plan comparison, checked 13 November 2026, speech and text analytics is included in CX 3 and CX 4 and an add-on for CX 1 and CX 2, where its features use Experience Units. Features such as topic mining and the agent, queue and flow Topics views list the WEM add-ons or CX 3 as their licence basis.

Transcription and the intelligence layer are billed separately. Transcription minutes come from the organisation's shared, pooled fair-use allocation whether or not a user has an add-on; the STA add-on is what unlocks the conversation intelligence that relies on them, such as topic spotting, sentiment, empathy and sensitive data masking, assigned with Billing > User > STA Upgrade. If someone without the add-on joins an interaction through a transfer or conference, the minutes they consume come from the pool, and usage beyond the pool is charged as overage. Widening transcription to more queues is therefore a coverage and cost decision, not just a technical one.

2.2 Permissions

You need the Speech and Text Analytics permissions for your role. Speech and Text Analytics > Data > View gives access to transcripts, topics and sentiment without granting recording access, which suits analysts. Permission changes take a few minutes to apply.

2.3 Where things are

• Menu > Conversation Intelligence > Speech and Text Analytics: Topic Miner, Topics, Programs and Speech and Text Configuration (Admin > Quality also reaches programs, topics and the miner).
• Menu > Analytics > Analytics Workspace: Topic Trends, and the agent, queue and flow Topics views.
• Performance > Content Search: search interactions by transcript content, sentiment and topics.

2.4 The real prerequisite: language and dialect

Transcription only works if the platform knows which language is arriving. Before you write a topic, confirm that voice transcription is enabled, that queues are set to transcribe, that the Transcription action is used in your flows, and that the language is right at the trunk and in Architect. Since December 2025 you can also set the transcription dialect at queue level, which matters most in multilingual centres and for outbound calls that never pass through a flow.

Two structural facts to design around: programs, topics and phrases are not available in every Genesys Cloud language, so check the supported languages before promising coverage for a site; and semantic topic spotting, generally available since May 2026, works in English only, so a multilingual estate runs semantic topics in English and lexical topics elsewhere.

4.1 Investigate in this order

1. Is the interaction transcribed at all, and in the right dialect? Transcription switched off for a queue, or the wrong dialect, produces a topic that never fires however good its phrases are.
2. Is the vocabulary being heard? Accuracy is measured as word error rate and is affected by audio and recording quality, connectivity, the engine, its training data, vocabulary, accents and speaking style.
3. Only then: are your phrases wrong?

4.2 Genesys's diagnosis method

Rule out systemic network, configuration or service problems; manually review and benchmark a sample of recordings; identify frequently mis-transcribed vocabulary; and verify that queues and flows have the right dialect. Genesys's stated best practice is to review at least two hours of randomly selected recordings across the organisation, and to ask supervisors or agents to log commonly mis-transcribed words during that review.

Then fix words in order: dictionary management for brand names, acronyms and internal terms; and, if that does not resolve it, data donation, where redacted recordings are donated to improve Genesys's native engine in a structured collaboration that takes several months.

4.3 Dictionary management

Use the dictionary when you have a manageable list of industry terms, when key terms are consistently misheard, split or dropped, and before launching products, services or campaigns with new names. There is one dictionary for the whole organisation; each term belongs to a single dialect and transcription engine; terms are case-insensitive, up to 30 characters, and up to 1,000 per dialect; and changes can take up to 30 minutes to take effect. It works in standard and low-latency transcription, and since April 2026 for Extended Voice Transcription Services dialects too.

5.1 Out-of-the-box topics

The out-of-the-box set is a starting point drawn from many organisations. Genesys is clear that you should adapt it to your own language, and that creating organisation-specific topics is a crucial step towards good results whatever you start from. Only one user can generate them at a time.

5.2 The default program and the IVR leg

If a default program is set and flows are mapped, the IVR portion of calls is topic-detected too. To stop that, set the default program to none, keep transcription as 'Enabled based on Queue configuration or Flow action', make sure no flows are mapped to programs, and publish. Decide this deliberately rather than discovering it in the data.

6.1 The limits that shape a mining plan

6.2 Designing the run

• Media type: you can mine calls, chat, email and messages. Mine one media type at a time: email phrasing is longer and more formal than chat, and mixing them dilutes the clusters.
• Participants: customers signal reasons and intent, agents signal process and compliance. Mine external participants for contact reasons, internal for mandatory wording, and both for a whole-journey view, expecting to filter agent boilerplate.
• Date range: long enough to cover a business cycle, remembering the 30,000-interaction ceiling. Record how many interactions were mined so the next run is comparable.
• Language: one per miner, because ignore lists are per dialect.
• Existing topics: leave the default exclusion on, so the miner skips phrases you already have and surfaces new patterns (configurable since July 2025, with a global exclude list).

6.3 Creating a miner, step by step

1. Go to Menu > Conversation Intelligence > Speech and Text Analytics > Topic Miner and click Create a New Miner.
2. Give it a descriptive name, for example date, queue group, media, dialect and purpose. In a year, the name is the only documentation you will have.
3. Choose whether to show or ignore existing topics and phrases, and select the language.
4. Under Data Source choose Genesys Cloud, then set the date range, media type, participants and up to 50 queues.
5. Confirm. Mining runs in the background; the status column shows progress, and a failed session can be corrected and restarted.
6. To re-run with different filters, duplicate an existing miner. To mine legacy recordings or another system's transcripts, choose External File and upload JSON in Genesys's example format.

6.4 Reading the results

For each mined topic you see the percentage and number of conversations that include it, the number of sentences (one turn by the agent or customer) and the number of phrase occurrences. Genesys's example: 'exchange rate' appeared in 21.74% of conversations, in five conversations and five sentences. The ratio between those numbers is your first quality signal: a high conversation count with one sentence each is a shallow, generic mention; several sentences per conversation is a real theme.

Click a topic to see its phrases, and use View Snippets to see the conversation fragment around a phrase; the miner masks personally identifiable data. Read the snippets before adopting anything: a phrase alone does not tell you whether the customer is asking for a cancellation or complaining about a cancellation fee, and that difference is the whole topic design.

6.5 Triaging a completed miner

This is the step people skip, and where the quality of the estate is decided.

1. Pass one, is it a business question? For each theme ask whether a named person acts on it, and mark it build, merge into an existing topic, ignore or park.
2. Pass two, read the evidence. For build and merge themes, open the snippets and check meaning (does the phrase mean what you think in your business?), specificity (would it also fire on conversations you do not want?) and speaker (customer, or an agent reading a script?).
3. Pass three, clean the noise. Add junk phrases to the phrase ignore list (by dialect and participant), and whole themes to the topic ignore list, which also ignores their phrases. Ignored items can be restored later.
4. Record the run: queues, dates, media, dialect, interactions mined, and the decisions you took and why.

7.1 Mined phrase to topic

From a completed miner, select phrases and either add them to an existing topic or click New Draft Topic, name it and add them. Then refine the topic in the Topics editor. Define topics before adding them to a program, add more phrases to make a topic more useful, and put double quotation marks around the word that carries the meaning, for example waive my "fee", so the phrase cannot match without it.

7.2 Topic to program, and program to queues and flows

A program is mapped to queues or flows and applies one set of rules to everything processed through them: topic detection, sentiment, program-level sentiment and empathy controls, AI summary and insights, and automated agent scoring. Open the program, select topics on the Selected topics tab, then save a draft or publish. Programs can contain topics in several languages and dialects. Genesys Cloud EX does not support flows, so there programs map to queues only.

7.3 Reusing a topic: share it or copy it

Sooner or later you want a topic that already exists, often in the out-of-the-box default program, in a new program. There are two different things you might mean. Sharing keeps one topic object in several programs: both search the same phrases at the same strictness, and one edit changes both. Copying creates an independent topic that starts with the same wording and can be tuned, published and retired separately. There is no move: a topic is not owned by a program, so to take it out of the old program, untick it there and republish.

Share by default: open the new program, tick the existing topic on the Selected topics tab, and save or publish. It uses no extra topic capacity and leaves one definition to maintain. Copy when phrases or strictness must differ, when the participant scope differs, when you want a clean data series from a known date, or when a topic feeds an automated score or regulated disclosure and needs a change freeze. Each copy uses one of your 500 topics (5,000 on request, an irreversible change) and must be tested on its own.

• The default program applies to all interactions not serviced by another active program, so a shared topic keeps firing there until queues are mapped to an active program.
• A topic is one object: editing it changes every program that includes it, including the default. Copy out-of-the-box topics before tuning a variant.
• Adding a topic to a program changes that program from the publish onwards, not history. To apply it backwards, reprocess within the existing program the history belongs to (see 13.3).

7.4 Publish, and mean it

Publishing is what changes production. An edited but unpublished topic, or a program not republished after a topic change, leaves detection running on the old definition while you believe your change is live.

7.5 Connect it to a consumer

A topic nobody uses is a cost. Each consumer is a distribution channel for your findings.

• Analytics views: Topic Trends, and the Topics tabs in agent, queue and flow views.
• Content Search: search by transcript content, sentiment and topics.
• Interaction categories: combine topics into situations (section 10).
• Process automation triggers: rules that start workflows based on topics, sentiment and empathy. Genesys's examples include notifying billing when a billing issue topic is detected and alerting a supervisor when sentiment falls below minus 20.
• Quality management: automated agent scoring rules in a program, up to 20 per program.

8.1 Choose the approach

• Compliance and disclosure wording is lexical: you need the exact words, at precision, and accept the maintenance.
• Intent and demand topics are semantic: you want the intent however it is said, at slightly lower precision and much lower maintenance.
• Semantic topics have tighter limits: up to 20 phrases per semantic topic and 50 semantic topics per dialect per program, against 1,000 phrases per lexical topic. English only for now.

8.2 Design the topic

Genesys's four design principles: start with one business objective; focus on customer intent; use customer language; create measurable topics. Combining several intents in one topic reduces accuracy and makes reporting difficult.

• One objective: write it in the topic description. If you cannot finish the sentence 'when this fires, we will...', the topic is not ready.
• Customer intent: name the topic for the intent, not the product or queue. 'Wants to cancel' is an intent; 'broadband' is a noun that will fire on a large share of calls.
• Customer language: take phrases from miner snippets, not imagination.
• Measurable: connect it to a metric someone owns, such as retention, first contact resolution, handle time for a contact reason or compliance coverage.

8.3 Build the phrases

A topic needs at least one phrase and can hold up to 1,000 (lexical); phrases can be added one at a time or imported in bulk, and the same phrase can appear in several topics. For lexical topics, build wording variations: vary verbs and openings, and quote the mandatory word. For semantic topics, build a small number of representative utterances that express the intent.

The Generate Phrases button suggests phrases with AI, up to 10 per request and 100 requests a day per organisation, without consuming Experience Units. Suggestions are not used for detection until you review, approve and save them, and Genesys says to review each one individually; the API marks them as AI-generated so you can audit them.

8.4 Strictness

Strictness is the precision and recall dial, set at topic and phrase level. Genesys's guidance is to start with the default, evaluate the captured events, then adjust, because experimentation is the way to balance precision and recall; longer phrases with more meaningful words often need less strict matching. In the API, strictness takes one of six values (1, 55, 65, 72, 85 and 90), so treat it as named steps rather than a slider. Phrases also carry a sentiment field in the API, which Genesys documents as not in use; sentiment comes from a separate model.

9.1 The validation gate

Build two sets of known interactions for each topic, around 20 that should be detected and 20 that should not, from miner snippets, Content Search and supervisors' examples, and reuse them every time you tune.

9.2 A tuning method that works

When a topic misbehaves, change one thing at a time, in this order.

1. Too few hits, and the words are missing from the transcript: transcription or dictionary problem (section 4).
2. Too few hits, and the transcript contains the words: add phrases from real snippets, or lower strictness one step.
3. Too many hits on one wrong phrase: raise that phrase's strictness, quote the mandatory word, or delete it.
4. Too many hits across phrases: the topic is too broad; split it, or narrow it to the customer side.
5. Fires in the wrong place: check the topic dialect and the program's queue and flow mappings.

11.1 How the scores are built

Sentiment analysis classifies each customer phrase as positive, negative or neutral from the language used, gives each a magnitude, places markers in the waveform and transcript, and derives an overall score and trend for the interaction. The overall customer sentiment score runs from minus 100 to plus 100 and weights events towards the end of the interaction, answering whether the customer left happy or dissatisfied; Genesys publishes the normalisation as 100 × tanh(0.75 × overall sentiment score).

Agent empathy classifies each agent phrase as empathetic, neutral or unhelpful and scores each agent per interaction from minus 100 to plus 100: 100 × (empathetic minus unhelpful phrases) ÷ (empathetic plus unhelpful phrases). Two empathetic and three unhelpful phrases give minus 20. Portions with two or more agents talking to the customer at once are excluded.

11.2 Where it is switched on

Organisation settings are the top level of control. Program-level settings can restrict sentiment and empathy further but cannot enable them if the organisation setting is off, which lets you handle regional, operational or compliance differences between programs; the Programs list shows each program's current state. Sentiment analysis is enabled automatically with voice transcription and can be switched on or off for the organisation in the STA settings.

11.3 Feeding the model back

Sentiment feedback is not required; sentiment works out of the box. It is for correcting phrases the model tags wrongly: you enter dialect-specific phrases as positive, negative or neutral, up to 2,000 statements per organisation, each 1 to 20 words. Work from the outliers: filter Content Search for strongly negative scores, read the transcript, and add feedback once for a mis-tag that recurs in your business, such as an industry idiom. Remember that the model uses surrounding context, so the same phrase can rightly score differently, and that poor audio causes missing or wrong markers; listen to interactions with many negative markers before concluding anything.

11.4 Compliance note for UK and EU estates

Genesys states that its sentiment analysis does not fall under the EU AI Act's restriction on emotion recognition in the workplace, because it analyses the language used and scores phrases without using biometric data, and that the same applies to agent empathy analysis, which evaluates the agent's language. That is Genesys's position; take your own legal view, but it is the fact to start from when representatives ask, and a reason to keep program-level controls tidy.

12.1 Topic Trends

The Topic Trends summary shows topics over time; click one for the detail view by interval, filtered by date, media type and interaction details, and save the view. Exports are limited to 1,000 records, and the view does not update itself: click Refresh. Digital topic data lags by days: STA needs 72 hours of inactivity to decide an email or messaging conversation has ended, metrics use the session end time, and topic reporting therefore updates 72 to 96 hours after the session ends.

12.2 Agent, queue and flow Topics views

The Topics tabs turn topic data into metrics by agent, queue and flow, alongside handle time and sentiment, and Genesys's stated use is to look for outliers between agents. Use them as a distribution test: a topic that looks sensible overall but fires on 90% of one agent's interactions and 1% of another's is telling you about a script, a mapping or attribution, not customer intent. Genesys suggests quality managers use the agent view for coaching and recognition, analysts look at trends by handle time to find contact reasons that make calls longer, and risk managers review key topics periodically with ad hoc searches.

12.3 Content Search

Content Search is your sampling engine: search transcript content (exact, similar or not similar), filter sentiment score or trend, include or exclude topics, combine terms with AND and OR, filter and show Program and Detected Category, sort by date, duration and sentiment, and save the view so the same audit query runs identically next month. Open a topic's matches, read 20 snippets and classify each as true or false: raise strictness or delete phrases for false matches, add phrases from interactions the topic missed, and record the precision.

12.4 Cross-check against something independent

1. Wrap-up codes: if agents code cancellation on 4% of calls and the topic says 22%, one of them is wrong.
2. Survey reasons and complaint records: do they agree with the topic or sentiment?
3. Recordings: re-run the two-hour random sample after significant dictionary or engine changes.
4. Handle time and resolution: a real contact-reason topic should have a plausible handle time.
5. Quality evaluations: where a topic feeds automated scoring, check that automated signals and human evaluators agree, in the same spirit as calibration.

13.1 The operating rhythm

13.2 Event-driven changes

13.3 Reprocessing history when a topic changes

Reprocess Interactions applies topic changes retrospectively, so new or changed topics are reflected in past conversations; before it, new topics only affected future interactions. It is available in the interface and the API, and Genesys currently provides it at no charge, with at least 30 days' notice if billing is introduced, so treat that as temporary.

• Scope is by program. A brand-new program is not included, because historical interactions are not linked to it: add new topics to the existing program the history belongs to.
• Up to 250,000 interactions per job, one job running at a time per organisation (with up to 10 queued), a start date up to 365 days back and an end date at least a day back.
• Interactions with no detected topics are marked failed, so a high failure count can simply mean nothing of interest was found.
• The CSV lists only newly detected topics, with no removed topics or before-and-after comparison.
• Genesys's best practice: reprocess only the programs tied to updated topics, use date and media filters to limit volume, and keep the exported interaction IDs for audit.
• Categories are never applied to history; reprocessing is for topics only.

13.4 Governance: register, naming, change control and people

None of this is a Genesys feature, and all of it is what lets an estate survive a change of owner. Keep a register with one row per topic, program, category, dictionary term, sentiment feedback statement and miner run: the business question it answers (if blank, retire it), owner and consumers, dialect, media and participant scope, matching type and strictness, programs, phrase sources (miner, AI-generated, supervisor, complaint), validation results with sample size, publish date, and next review.

Name topics in business language for the intent, and programs after the outcome owner, never a queue. Use tags sparingly as review filters. Require a second person to review anything that feeds a score. Check a topic's programs before deleting it. Publish in a standing window. Use the audit log to see who changed what. Where you have a non-production organisation, build and tune there and promote deliberately, remembering that STA objects move by API or re-entry and the dictionary is per organisation.

Name the owner. A medium-sized estate needs one owner for topics, programs, dictionary and categories; supervisors and quality evaluators as feeders who report wrong transcripts and detections; an analyst as consumer; and a business sponsor per program to decide what is worth measuring. Protect around half a day a week for the rhythm above, because an unowned estate quietly becomes an untrusted one.

14.1 Best practice in one place

1. Lexical for compliance wording, semantic for intent, and never the same phrase strategy for both.
2. One business objective per topic, written in the description, in customer language, and measurable.
3. Adapt out-of-the-box topics; organisation-specific topics are what make STA work.
4. Program by outcome owner, not by queue.
5. Share a topic when the definition should stay identical; copy it when phrases, strictness, scope or a change freeze differ.
6. Run a use-case workshop before a miner; mine one media type and one dialect at a time.
7. Read the snippets before adopting a phrase, and curate ignore lists every run.
8. Test phrases against history before publishing; move strictness one step at a time.
9. Review AI-generated phrases one by one; nothing detects until approved.
10. Measure precision monthly with a stated sample size, and cross-check against independent data.
11. Fix transcription before blaming a topic; the two-hour random sample is the baseline.
12. State the 72 to 96 hour lag on digital topic reports.
13. Reprocess only within existing programs, narrowly, and announce restated trends.
14. Retire topics nobody uses, and treat an approaching limit as a design signal.
15. Tell people whenever the definition of something they rely on changes.

14.2 Troubleshooting

14.3 Limits reference

From Genesys's STA articles and the Developer Center limits, checked 13 November 2026. Check them again before designing near a limit.

14.4 API for automation and audit

Everything in the STA editors is available through the Speech and Text Analytics API under /api/v2/speechandtextanalytics, useful for bulk change, promotion between organisations and automated checks. Useful endpoints include /programs and /programs/mappings for inventory; /programs/publishjobs and /topics/publishjobs for publishing; /topics and /topics/{topicId}, which return strictness, matching type, participants, dialect, phrases, programs and published state; /topics/testphrase/jobs for phrase tests; /topics/generatedphrases/jobs for AI phrase suggestions; and /sentimentfeedback. For lower-latency transcripts, subscribe through the Notifications API.