fonnx models

ONNX models for fonnx, which runs ONNX models in Flutter on every platform.

fonnx pins each file by the commit of this repository and its SHA-256 (see native_artifacts/manifest.json in fonnx). Do not squash the history of this repository or delete files from it: shipped apps download files by commit.

Each file has the license of its source model.

Files that fonnx pins

File	Source	License
decoder-epoch-12-avg-2-chunk-16-left-64.int8.onnx, encoder-…, joiner-…	sherpa-onnx keyword spotter, GigaSpeech Zipformer 3M	Apache-2.0
fastconformer_ctc_multilingual_int8.onnx	NVIDIA stt_multilingual_fastconformer_hybrid_large_pc, CTC branch, int8 export of sherpa-onnx	CC-BY-4.0
futo_swipe_*.onnx	Derived from FUTO Swipe (see below)	FUTO Model Weights License 1.0
kokoro_prosody.onnx, kokoro_decoder.onnx	hexgrad/Kokoro-82M v1.0, uint8 export of onnx-community/Kokoro-82M-v1.0-ONNX at 1939ad2a8e416c0acfeecc08a694d14ef25f2231, split in two by fonnx	Apache-2.0
kokoro_voice_*.bin	The 28 English voices of the same onnx-community revision, unchanged	Apache-2.0
kokoro_misaki_{us,gb}_{gold,silver}.json	Pronunciation dictionaries of hexgrad/misaki 0.9.4, unchanged	Apache-2.0
kokoro_g2p_fallback_en_us.onnx, ….json	PeterReid/graphemes_to_phonemes_en_us at a5631b285d18d59483c32c0c3379cb9fac924f4b, ONNX export by fonnx	Apache-2.0
kokoro_spacy_en_core_web_sm.json, …_tagger.bin	spaCy en_core_web_sm 3.8.0, tokenizer rules and tagger weights converted by fonnx	MIT
magika.onnx	Google Magika	Apache-2.0
miniLmL6V2.onnx	sentence-transformers/all-MiniLM-L6-v2	Apache-2.0
msmarcoMiniLmL6V3.onnx	sentence-transformers/msmarco-MiniLM-L-6-v3	Apache-2.0
parakeet_tdt_v3_decoder_int8.onnx, …_encoder_…, …_joiner_…	nvidia/parakeet-tdt-0.6b-v3, int8 export of sherpa-onnx	CC-BY-4.0
potion8m.onnx, potion32m.onnx	Minish Lab potion-base-8M and potion-base-32M	MIT
pyannote_seg3.onnx	pyannote/segmentation-3.0	MIT
sense_voice_small_int8.onnx, sense_voice_small_tokens.txt	FunAudioLLM/SenseVoiceSmall (Alibaba), int8 export of sherpa-onnx (2024-07-17)	FunASR Model License
silero_vad_v6.2.1.onnx	Silero VAD v6.2.1	MIT
whisper_tiny_decoder.onnx, whisper_tiny_encoder.onnx	onnx-community/whisper-tiny, from openai/whisper-tiny	MIT

Other files

whisper_{base,small}_{encoder,decoder}.onnx are from onnx-community/whisper-base and onnx-community/whisper-small (MIT). bpe_decoder.onnx, silero_vad.onnx, whisper_{tiny,base,small}.onnx and whisper_v3_turbo_olive.onnx are files of earlier fonnx versions.

Attribution

The NeMo ASR files are exports of NVIDIA models, used under CC-BY-4.0. fonnx did not change the files of the sherpa-onnx exports.

The Kokoro files are from Kokoro-82M by hexgrad, used under Apache-2.0. fonnx split the onnx-community export into a prosody model and a decoder model; the weights are unchanged. The misaki dictionaries, the fallback model and the spaCy data are used under their licenses above.

The SenseVoice files are the sherpa-onnx export of SenseVoice Small by Tongyi Lab, Alibaba, unchanged, used under the FunASR Model License.

The futo_swipe_*.onnx files are derived from FUTO Swipe and are subject to the FUTO Model Weights License 1.0. LICENSE-FUTO.txt in this repository has the complete, unchanged terms. They are an unofficial ONNX conversion; this does not imply sponsorship, affiliation, or endorsement by FUTO. No training or intentional weight changes were made. A product that uses them must show its users that it is powered by "FUTO Swipe" technology.