CoolFace
Modelpublic

mira-thing/mira-voice

sourceHugging Faceotherupdated 3mo agoView on Hugging Face
0likes56downloads
Model Card

mira-voice - prebuilt on-device voice runtime (aarch64)

Prebuilt model + runtime bundle for [Mira](https://github.com/mira-thing/mira-voice), the 100%-on-device smart-voice stack for the discontinued Spotify Car Thing. This repo holds the heavy, upstream artifacts. The code, build scripts, and the one custom model (the "hey mira" wake word) live on GitHub at [mira-thing/mira-voice](https://github.com/mira-thing/mira-voice).

This is a deployment bundle, not a transformers loadable model. Fetch it via the repo's fetch-artifacts.sh, or huggingface-cli download mira-thing/mira-voice.

Contents

PathWhatUpstreamLicense
zipformer/...gigaspeech.../sherpa-onnx Zipformer int8 ASR (encoder/decoder/joiner + tokens/bpe)k2-fsaApache-2.0
lib/aarch64 runtime: ONNX Runtime, TFLite, sherpa-onnx, glibc family, audio libsmixedsee THIRD_PARTY_LICENSES
models/melspectrogram.tflite, embedding_model.tfliteopenWakeWord shared base modelsopenWakeWordApache-2.0
espeak-ng-data/espeak-ng g2p data, trimmed to en-usespeak-ngGPL-3.0
bin/aarch64 helpers: oww_wake, sherpa_asr_server, espeak-ngthis project + espeak-ngsee repo

Not in this bundle: the `hey_mira.tflite` wake model, version-controlled in the GitHub repo.

Use

sh
git clone https://github.com/mira-thing/mira-voice && cd mira-voice
bash fetch-artifacts.sh

The sherpa Zipformer is the unmodified k2-fsa GigaSpeech release; the runtime .so are upstream libraries cross-built for aarch64. Full attribution in `THIRD_PARTY_LICENSES`.