HuggingFace

We have safetensors versions of the models linked from our organization on HF, huggingface.co/moonshine-ai/models. These are all floating-point checkpoints exported from our training pipeline. The organization was called UsefulSensors until recently, after an earlier incarnation of the company when we were focused on supplying complete voice interface solutions integrated onto a low-cost chip with a built-in microphone; links under the old name no longer resolve.

Separately, every asset this library can download is mirrored in a single repository at huggingface.co/moonshine-ai/moonshine-voice-assets. It holds exactly the files the download manifests reference (currently 438 files, 7.3GB) — the .ort speech to text and embedding models, the TTS voices, and the grapheme to phoneme data — laid out in the same folder hierarchy as the CDN, so a path in that repo is the path under download.moonshine.ai. The list is enumerated from the compiled model catalog rather than by listing the bucket, because the bucket also holds unconverted .onnx sources, older copies of models that manifests no longer point at, and platform artifacts like the Raspberry Pi disk image. If you need to know which files a given release actually uses, that repo is the authoritative answer, and its FILES.tsv records the size and checksum of each one. Note that the library still downloads from the CDN at runtime; the mirror is there for archival and verification, not for serving. For local tests and offline TTS work, populate the gitignored trees with scripts/fetch-voice-assets.sh. Reclaiming GitHub LFS storage after removing historical binaries is documented in docs/lfs-purge.md.