moonshine-voice-assets

Moonshine Voice assets

Every file that the Moonshine Voice SDKs can download, in one place, with the CDN's folder hierarchy preserved. This repo exists to be an unambiguous source of truth: the serving bucket behind download.moonshine.ai also holds unconverted sources, superseded copies at older paths, and unrelated build artifacts, so "what does the library actually use?" was not answerable by listing the bucket.

438 files, 7.26 GB. Paths here match CDN paths exactly, so model/base-ar/quantized/base-ar/encoder_model.ort in this repo is https://download.moonshine.ai/model/base-ar/quantized/base-ar/encoder_model.ort.

This is a mirror for archival and verification. The SDKs still download from download.moonshine.ai at runtime, and that CDN remains the endpoint to point tooling at — nothing here changes how the library fetches models.

What's included

Two inventory files sit at the repo root:

• FILES.tsv — path, size in bytes, and the md5 and/or CRC32C the CDN reported for each file. This is the checked-in definition of the set.
• TRANSFER_REPORT.tsv — path, size, and sha256 computed while each file was copied in, so the Hub copy can be verified independently of the CDN.

How the file list was chosen

The set is not a hand-curated list or a glob over the bucket. It is enumerated from the compiled model catalog through the C API — the same source the SDKs themselves consult — by walking every combination a caller could ask for:

• every language and architecture in the STT catalog, requesting include_spelling and word_timestamps so the optional groups are covered
• every embedding model and quantization variant
• the diarization dependencies
• the TTS dependencies for every catalog language, once per voice available in that language

Every file URL that appears in any resulting manifest is included. Because the catalog is the source, this cannot drift from what the library downloads: a file the manifests no longer reference drops out, and a newly referenced file appears.

What's deliberately excluded

The serving bucket holds 801 objects; the 363 not listed here are excluded because no manifest references them:

• Unconverted ONNX sources (.onnx, .onnx_data) — the library loads the memory-mappable OnnxRuntime .ort conversions, so the upstream ONNX files are build inputs rather than shipped assets. Note that Piper's *.onnx.json voice configs are included: despite the name they are small JSON config files, not ONNX models, and the voices do not load without them.
• Superseded copies at older paths. Several models exist at more than one location, for example the manifests reference model/base-ar/quantized/base-ar/ while flat copies remain at model/base-ar/. Only the referenced path is here, which is much of the point of this repo.
• Platform artifacts: the Raspberry Pi disk image, Android APKs, desktop installers, and loose build scripts and docs.

Licensing

The assets here carry the licenses of their upstreams and are not under a single license. In particular:

• The English speech-to-text models are MIT-licensed.
• Models for other languages are released under the Moonshine Community License, which is non-commercial.
• TTS voices and grapheme-to-phoneme data follow the terms of the projects they come from (Kokoro, Piper, and the per-language dictionary sources), and EmbeddingGemma is subject to Google's Gemma terms.

See the License section of the main README and core/moonshine-tts/data/README.md for the per-language and per-voice detail. Check the terms that apply to the specific files you intend to use before redistributing or building on them.

Getting the files

Normal use needs none of this — the SDKs fetch what they need on demand. To pull the archive directly:

# everything
hf download moonshine-ai/moonshine-voice-assets --local-dir moonshine-assets

# just one model
hf download moonshine-ai/moonshine-voice-assets \
  --include "model/base-en/**" --local-dir moonshine-assets