Public Surface Inventory
This inventory defines the documentation scope for the current package. It is
an architectural map, not a promise that every implementation module is a
directly supported import path. The machine-readable source of truth is
docs/namespace_inventory.json; every listed namespace has a support level,
landing/reference page, and owner page. Run
python docs/check_namespace_coverage.py to validate it.
Stable package areas
hyperion.ioAudio, feature, HDF5, ARK, Kaldi-style, and VAD readers/writers. The public reference will cover reader/writer contracts, supported formats, key and dtype semantics, and factory behavior.
hyperion.utilsDataset manifests,
HyperDataset, Kaldi-style metadata tables, enrollment mappings, trial indexes/keys/scores, and supporting utilities. These types form the package’s core data model.hyperion.metricsandhyperion.np.metricsHigh-level evaluators and the underlying verification/scoring metrics.
hyperion.data_prepThe
DataPrepextension point and maintained dataset preparation implementations.hyperion.npNumPy models, PDFs and PLDA, classifiers, transforms, calibration, score normalization, preprocessing, clustering, diarization, augmentation, and feature utilities.
hyperion.torchThe core PyTorch stack: primitive layers, layer blocks, neural architectures, models, datasets/samplers, trainers, losses, optimizers, schedulers, loggers, metrics, transforms, and shared utilities. Curated reference pages will prioritize documented package exports and model/trainer extension points over every implementation module.
hyperion.torch.tpmStable integrations with external model and evaluation packages, including Hugging Face, DNSMOS, UTMOS, and VoxProfile. Their pages must identify optional dependencies and external model requirements.
hyperion.torch.adv_attacksandhyperion.torch.adv_defensesStable adversarial attack and defense interfaces, factories, and supported evaluation workflows.
hyperion.text_normText-normalization base classes and English normalizers.
hyperion.binAll command modules that produce installed
hyperion-*entry points. The CLI reference will classify commands by task and document their parser, inputs, outputs, and configuration schema.
Experimental package areas
hyperion.torch.models.dacNeural codec models and streaming codec behavior.
hyperion.torch.models.freevcand VI anonymizer workflowsVITS-based anonymization and voice-conversion interfaces.
hyperion.torch.models.transducerandhyperion.torch.models.wav2transducerTransducer architectures and waveform-to-transducer wrappers.
hyperion.torch.models.qvectorsQ-vector models, wrappers, trainers, and inference workflows.
The associated commands remain part of hyperion.bin but will be labelled
experimental in the CLI reference.
Out of scope
egs/recipes and their shell/tooling dependencies.hyperion.bin_deprecandhyperion.bin_deprec2(hyperion/bin_deprec*/on disk).hyperion.helpers: compatibility-oriented composition helpers that are currently internal rather than a supported extension surface.Private names and incidental implementation modules that are neither exported nor required as documented extension points.
Inventory maintenance
When adding a package area, command, or top-level extension point, update this
page, docs/namespace_inventory.json, and Documentation Policy in
the same pull request if its support level or scope changes.