nanobot-channel-voice-models

nanobot-channel-voice models

RKNN model artifacts and a checksum-pinned weights index for nanobot-channel-voice.

Install

The repository includes four official openWakeWord phrase heads, each available as an original-ONNX package for CPU/ONNX-capable hosts and as a validated RV1126B hybrid package:

wake/openwakeword/alexa/onnx
wake/openwakeword/alexa/rknn.rv1126b
wake/openwakeword/hey-jarvis/onnx
wake/openwakeword/hey-jarvis/rknn.rv1126b
wake/openwakeword/hey-mycroft/onnx
wake/openwakeword/hey-mycroft/rknn.rv1126b
wake/openwakeword/hey-rhasspy/onnx
wake/openwakeword/hey-rhasspy/rknn.rv1126b

Choose one variant for the target host. Every fetch is a complete package: ONNX entries install the three original upstream graphs plus a phrase sidecar; RV1126B entries install the phrase head, exact Python-mel filterbank, and static RKNN embedding. The runtime fetches a complete model entry through nanobot-voice:

nanobot-voice list
nanobot-voice fetch stt/whisper/base/rknn.rv1126b --yes

To use this index explicitly:

export NANOBOT_VOICE_INDEX=https://huggingface.co/iChizer0/nanobot-channel-voice-models/resolve/main/weights-index.json
nanobot-voice list

weights-index.json is the authoritative manifest of available keys, source provenance, checksums, and acceptance notices. Files under models/*/_shared/ are implementation storage for shared dependencies; they are not selectable model keys.

License and redistribution notices

Each model entry has its own license and provenance in weights-index.json. Review and comply with those terms before use or redistribution. In particular:

• The openWakeWord entries (both .../onnx and .../rknn.rv1126b) contain official pretrained phrase heads, which are CC-BY-NC-SA-4.0. They are non-commercial and share-alike; these are not unrestricted wake-word models. ONNX entries use the original v0.5.1 melspectrogram.onnx and embedding_model.onnx backbone. RV1126B entries use the exact Python mel frontend plus static fp16 RKNN embedding, while each small phrase head remains ONNX.
• The converted melspectrogram.rknn is deliberately not distributed because its RV1126B numerical output is incorrect. The device-neutral mel_filters.npy is the frozen upstream 257x32 matrix required by the validated Python frontend.
• The openWakeWord package does not include the upstream test WAV or any unverified third-party frontend data.

The repository's four openWakeWord packages are documented in OPENWAKEWORD_WEIGHTS_LAYOUT.md, which records the hybrid runtime contract, portable phrase-head storage, physical shared-file layout, selected checksums, and post-upload validation procedure.

• MMS-TTS English is CC-BY-NC-4.0 (non-commercial use only).
• DataBaker Chinese Matcha (tts/matcha/zh-CN/baker/...) is subject to the DataBaker Chinese Standard Mandarin Speech Corpus non-commercial terms.
• The LJSpeech Matcha entry requires review of the LJSpeech Dataset License and applicable upstream Matcha, Vocos, and sherpa-onnx terms.
• The bilingual tts/matcha/zh-en/icefall/... entry is a derived RV1126B fp16 static split of csukuangfj/matcha-icefall-zh-en and its matched 16 kHz Vocos. The ModelScope acoustic source reports Apache-2.0 and Vocos is MIT. The entry ships the upstream espeak-ng-data English frontend pack as espeak-ng-data.tar.bz2, which nanobot-voice fetch unpacks beside the model files (the file is marked "extract": true in the index). The pack is byte-identical to the upstream model repository's espeak-ng-data/ (355/355 files sha256-verified) and to the pack used by the on-device validation harness — it is the training pairing, so no espeak-ng version pinning is needed. espeak-ng is GPL-3.0; the license text and provenance note ship as espeak-ng-data.LICENSE.txt. This entry requires a nanobot-channel-voice version with the static zh-en-lexicon Matcha frontend and eSpeak en-us support for English; a plugin too old to extract archives installs the archive as a plain file instead — unpack it beside the model files or point tts.matcha.espeakDataDir at an extracted copy.

Index stability

• Published files are immutable: bytes never change under the same URL. A content change ships under a new file name or a new key, so a sha256 mismatch on fetch means corruption or a stale cache — never a silent replacement.
• Key renames keep a deprecated alias (deprecated: true, renamed_to) in weights-index.json for two index generations. See MIGRATION.md for the rename table and mechanical config-rewrite instructions.

No license is granted here beyond the licenses and terms of the individual upstream artifacts.