Experimental Components

The components on this page are part of the maintained package, but they do not carry the compatibility guarantees of the stable APIs. Their Python interfaces, configuration schema, training semantics, checkpoint layout, and output behavior may change in a minor release. Pin the Hyperion revision and preserve the full configuration, dependency versions, and a small inference regression test with every trained artifact.

This page intentionally does not provide recipe instructions from egs/. Use the installed command’s --help output and a version-controlled jsonargparse configuration as the source of truth for the version you run.

Neural codecs (DAC)

The neural codec family is implemented in hyperion.torch.models.dac and uses DACTrainer. Both standard and streaming variants are experimental.

  • hyperion-train-dac

  • hyperion-finetune-dac

These commands use the current AudioDataset and SegSamplerFactory configuration contract. Training can require a discriminator and audio-feature loss configuration in addition to the codec model. Treat encoded streams and decoder checkpoints as version-coupled: verify encode/decode behavior after an upgrade before relying on saved artifacts.

VITS anonymization and voice conversion

hyperion.torch.models.freevc provides FreeVC and Hugging Face WavLM-based voice-conversion models. VI anonymizer workflows build on this family and an audio discriminator; their trainers are FreeVCTrainer and VIAnonymizerTrainer.

  • hyperion-train-freevc

  • hyperion-train-vi-anonymizer

  • hyperion-finetune-vi-anonymizer

These workflows may require externally sourced pretrained models and substantial GPU memory. Keep the pretrained model identifier or local revision, sample frequency, speaker-conditioning setup, and privacy/quality evaluation protocol beside each checkpoint. Anonymization claims require task-specific evaluation; a successful training run alone is not evidence of privacy protection.

Transducers

The transducer families are in hyperion.torch.models.transducer and hyperion.torch.models.wav2transducer. Their training code uses TransducerTrainer. Some maintained paths depend on k2 and text/token metadata supplied through the audio dataset configuration.

  • hyperion-train-wav2rnn-transducer

  • hyperion-train-wav2vec2rnn-transducer

  • hyperion-train-wav2vec2transducer

  • hyperion-finetune-wav2vec2transducer

Record the tokenizer model and vocabulary, blank/special-token convention, decoder settings, and exact k2 version with every checkpoint. The decoder commands are available for inspection but have not yet received a maintained task tutorial; validate their input and output contract on fixture-scale audio before a full run.

Q-vectors

Q-vector models and wrappers are located in hyperion.torch.models.qvectors and use QVectorTrainer. They are not interchangeable with the stable x-vector extraction and scoring interfaces.

  • hyperion-train-qvector

  • hyperion-finetune-qvector

  • hyperion-infer-qvectors

Version the quantizer/head configuration and all upstream acoustic or Hugging Face model revisions with the checkpoint. Check output tensor layout and the meaning of inferred codes for the selected model rather than assuming an x-vector-compatible embedding matrix.

Adoption checklist

Before using an experimental component in a long-lived system:

  • pin the Hyperion commit or released package version;

  • retain the complete YAML/JSON configuration and optional dependency versions;

  • save a known input/output regression fixture alongside the model;

  • test saving, loading, and inference after every upgrade; and

  • explicitly validate task outcomes such as recognition, intelligibility, codec fidelity, or anonymization privacy.

See also