helena 0.1.0

Core types and component interfaces for helena, a latent data-to-waveform generation platform.
Documentation

helena

Core types and component interfaces for helena, a latent data-to-waveform generation platform: it learns to turn structured source data into audio by predicting compressed audio latents that a frozen neural codec decodes.

This crate is the foundation the rest of the platform builds on. It holds the vocabulary every other crate speaks — the data structures that flow through the pipeline and the traits the swappable components implement — and nothing that pulls in a tensor backend, a neural network, or an FFT. It is host-side, serde-friendly, and dependency-light by design.

source data ──▶ DataEncoder ──▶ ConditioningLatents ──▶ ConditionalGenerator
                                                             │
                                                      AudioLatent / tokens
                                                             │
                                                         AudioCodec::decode
                                                             ▼
                                                          Waveform

What's here

  • Pipeline data — Tensor (a host-side, serde-friendly wrapper over Burn's TensorData), Waveform, the conditioning and audio latents (ConditioningLatents, AudioLatent), and the experiment/generation config and manifest records.
  • Component seams — the three traits the platform is built to swap parts behind: DataEncoder, AudioCodec, and ConditionalGenerator. They are backend-agnostic: a model's tensor backend never leaks into its callers.
  • The diffusion kernel — variance-preserving v-prediction levels, a cosine schedule, classifier-free-guidance scheduling, and Min-SNR loss weighting.
  • Subsystems behind their modules — generator-output diagnostics under eval, streaming statistics under stats, content hashing under hash. The crate root stays the vocabulary; these are reached through their module path.

Backend

The tensor backend is Burn. helena stays backend-agnostic by speaking Burn's host-side TensorData at its boundaries, so latents serialize, hash, and disk-cache without a device. Models in the (currently unpublished) training and inference crates are generic over B: Backend and convert to Tensor<B, D> at their edges.

Status

Pre-1.0 and evolving; the public API may change between 0.x releases. The other helena crates (data, model, train, infer, eval, cli) are not yet published.

License

AGPL-3.0-only.