Contact center quality assurance, in the detail your evaluation needs.

02 / Hybrid sentiment

Call1 scores the words and the voice separately, then fuses them.

Every eligible, speaker-aligned chunk can produce two independent affect signals: textual sentiment from what was said and tonal emotion from how it was said. A policy-versioned fusion step turns them into one useful hybrid result.

Both source signals stay attached to the result. If the language reads neutral while the voice carries strain, your reviewer sees the disagreement. And when audio quality is too poor for reliable tonal inference, the tone signal is marked unavailable.

Analysis unit

Each result attaches to a speaker-aligned chunk with start/end timestamps, speaker role, transcript text, and source quality indicators.

Independent outputs

Text and tone models retain their own label distribution, confidence, model identifier, model version, and inference time. Neither model substitutes for the other.

Fusion policy

The hybrid label, confidence, and disagreement score are produced by a separately versioned policy. Reviewers can inspect the contributing signals behind the fused result.

Acoustic context

The chunk schema can also carry interruption, silence, crosstalk, hold music, static, DTMF, or customer-defined event tags without retaining the source audio indefinitely.

04 / Smart review queue

The queue turns total coverage into focused human work.

Scoring every call should not create a larger pile for reviewers. Call1 ranks the conversations and moments that deserve attention: compliance failures, low-confidence decisions, emotional escalations, rare events, and unusual changes in performance.

Reviewers open a prepared case with audio, transcript, evidence, and the relevant scorecard question already in context. They confirm or override with a reason code and move on. Reviewers pull the next call; nobody cherry-picks an easy one.

Under-buy your capacity block and the queue clears slower. Nothing goes unscored, but the wait grows until you add a block.

Queue policy

Priority can combine compliance severity, failed questions, confidence thresholds, emotion changes, tracked events, age, and customer-defined campaign or team rules.

Review workspace

The waveform, speaker-aligned transcript, flagged timestamps, question rail, evidence spans, and playback position stay synchronized. Keyboard controls support play, jump, confirm, override, escalate, and submit-next.

Human decision record

Overrides require a reason code and may include a note. The original machine score stays intact; the reviewer's action is appended with user, timestamp, and provenance.

Operating roles

Reviewer, QA lead, and administrator responsibilities separate adjudication, queue management, calibration, reporting, rubric administration, and access management.

06 / Versioned intelligence

Change models without rewriting history.

Speech, text, tone, embedding, and scoring models will improve. Call1 treats each one as a replaceable, versioned component behind a stable per-chunk data contract, so the product can adopt better models without rebuilding the review experience.

New candidates can run beside the current model and compete on the same calls. Champion and challenger outputs keep their own confidence, version, and provenance. A swap changes future processing and leaves the historical record as it was.

Registry and lineage

Speech, embedding, text, tone, fusion, and QA components have stable identifiers and immutable versions. Every output states which component and policy produced it.

Evaluation modes

A production champion and one or more challengers may process the same eligible chunks. Shadow results remain isolated from customer-facing scores until an approved promotion.

Promotion and rollback

Promotion criteria are based on a customer-approved benchmark and defined acceptance metrics. Rollback changes the active version; neither action mutates already-published results.

Local improvement

Reviewer adjudications form a customer-owned labeled set for in-place evaluation or adapter training. Adapters trained on sensitive content stay within the deployment unless an approved sanitization and export policy exists.