Hyperion Data Model

Hyperion uses a small set of aligned tables to represent speech data and speaker-verification evaluation. Understanding these tables is more important than memorizing a particular command: the same model works for NumPy APIs, PyTorch workflows, and command-line tools.

The three levels of data

Media and segment manifests

RecordingSet describes source media. SegmentSet describes the speech segments, utterances, or clips used by a workflow. Other manifests attach data to those segment identifiers, including FeatureSet, VADSet, and DiarizationSet.

Dataset container

HyperDataset groups related manifests, class metadata, enrollment maps, and trial information. It is useful when a workflow needs to load, validate, filter, or save a complete dataset consistently.

Verification evaluation

Verification uses a matrix convention:

  • rows are enrollment model identifiers;

  • columns are test segment identifiers;

  • each cell is one (model, segment) trial.

TrialNdx marks trials to score. TrialKey adds the target/non-target ground truth. TrialScores stores system scores on the same grid. This alignment is essential: score matrices must use the same model and segment order as their key or index.

Typical flow

recordings -> segments -> features / VAD / labels
                           |
                           v
                    embedding extraction
                           |
enrollment map -> trial index/key -> scores -> metrics

An embedding system may supply the vectors from precomputed features or raw audio. The evaluation objects are independent of how those vectors were created.

Detailed guides