Build a private test set around the callers, channels, languages, and failure modes your voice agent actually faces. Spirelight scopes the collection, reference transcripts, slices, and holdout rules with your team.
A useful voice-agent evaluation set separates accent, language, channel, noise, interruption, and scenario effects. The result is not just one word error rate. It shows which caller groups and conditions cause the system to fail.
The collection is scoped around your deployment. If a hosted model cannot be fine-tuned, the same held-out set can still compare providers and detect regressions after each release.
Training leakage makes a benchmark look better without making the production system better. We define the evaluation split, file manifest, access rules, and any non-reuse commitment in the statement of work before recording starts.
Use the free worksheet to define coverage, rights, delivery fields, held-out rules, and acceptance tests before requesting a collection plan.
Build a dataset specificationSubmit your language and channel requirements. The team reviews the request manually and confirms whether an applicable conversational or telephony example is currently available.
Enter your work email and verify it with a 6-digit code. An admin reviews the request and confirms whether an applicable example is currently available and may be shared. Use the optional field to name the language, channel, or condition you need to inspect.
Send the language, channel, volume, annotation, and timing you know. The quote form opens directly in project-brief mode.
