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.io

Audio, 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.utils

Dataset 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.metrics and hyperion.np.metrics

High-level evaluators and the underlying verification/scoring metrics.

hyperion.data_prep

The DataPrep extension point and maintained dataset preparation implementations.

hyperion.np

NumPy models, PDFs and PLDA, classifiers, transforms, calibration, score normalization, preprocessing, clustering, diarization, augmentation, and feature utilities.

hyperion.torch

The 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.tpm

Stable 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_attacks and hyperion.torch.adv_defenses

Stable adversarial attack and defense interfaces, factories, and supported evaluation workflows.

hyperion.text_norm

Text-normalization base classes and English normalizers.

hyperion.bin

All 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.dac

Neural codec models and streaming codec behavior.

hyperion.torch.models.freevc and VI anonymizer workflows

VITS-based anonymization and voice-conversion interfaces.

hyperion.torch.models.transducer and hyperion.torch.models.wav2transducer

Transducer architectures and waveform-to-transducer wrappers.

hyperion.torch.models.qvectors

Q-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_deprec and hyperion.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.

See also