Unit testing for voice agents is more than checking if an intent fires. You need clear insight into how your NLU behaves under real speech conditions, where ASR errors, accents, background noise, and multi-turn logic all affect performance.
Cekura provides a purpose-built unit testing system that helps teams validate intent classification and entity extraction with the same rigor they expect from traditional software tests.
Test intent and entity accuracy with real conversational depth
Cekura gives you a full diagnostic view of how your agent interprets user inputs. Each test surfaces measurable signals that show exactly where the NLU succeeds or breaks down, including:
-
Intent correctness and misclassification patterns
-
Entity extraction accuracy including correctness, presence and span behavior
-
Per-entity precision and recall
-
Confidence score trends that reveal borderline cases
-
Turn-level timestamps highlighting where errors appear
This level of detail lets teams pinpoint the root cause rather than guessing why a conversation failed.
Create tests with simple structured cases or auto-generated scenarios
Teams can define unit tests in a declarative style. You can reference intents, entities and sample utterances directly or generate template sets automatically based on your agent description. Cekura also supports:
-
Parameterized tests to validate an intent across many slot values
-
Test sets tied to specific domains or workflows
-
Auto-suggested scenarios generated from your existing agent prompt or knowledge base
This gives you a low-maintenance testing workflow that scales as your NLU model evolves.
Ensure complete intent and entity coverage
Cekura makes it easier to evaluate your NLU’s completeness. Tests are organized for full coverage across intents, required entity types and edge-case utterances. You can include:
-
Negation and contradictory phrasing
-
Multi-intent or ambiguous inputs
-
Noisy transcripts
-
Accents and varied speaking patterns via predefined or custom personas
Tests can also highlight missing coverage gaps so nothing critical is overlooked.
Voice specific validation using ASR-aware evaluation
Since voice agents rely on ASR, Cekura checks the NLU against real ASR output instead of clean text. The platform can simulate common ASR mistakes, accent shifts and acoustic variations. This ensures your intent and entity unit tests reflect real call conditions.
Fits directly into CI workflows
Cekura provides REST APIs and CLI hooks so you can run your unit test suites automatically in CI. Tests can run on each model update, infrastructure change or prompt adjustment. Teams can also tag tests by feature area or intent group to control what runs for each pipeline.
Fast, scalable test execution
Cekura executes large batches of test utterances quickly and supports parallel execution. Every run captures versioned results, regression diffs and comparisons across builds. This gives teams a clear history of how intent and entity accuracy is trending over time.
Reporting that highlights exactly what matters
Each unit test produces an actionable summary with:
-
Pass and fail breakdowns
-
Intent confusion insights
-
Entity-level scoring
-
Time-stamped failure points
-
Exportable JSON, CSV, and PDF reports
Teams can drill into each utterance to see the transcript, metrics and explanation.
Automated drift and regression detection
Whenever a new model or prompt update changes NLU behavior, Cekura flags drops in intent accuracy, entity correctness or confidence thresholds. Alerts help teams catch regressions early without manually spot checking behavior.
Flexible for custom entity types and workflows
If you use custom slot types or domain specific entities, Cekura supports custom evaluators and input transformations. You can plug in your own logic or run tests under simulated context to reflect actual workflow state.