Third-party notices · onespeech docs

Model weights

Core embeds the Silero VAD v6.2.3 ONNX release model. No other model weights are included.

Data

The LibriSpeech clip identified below is the only third-party audio fixture; ID-only LibriSpeech evaluation lists are also included, with attribution below. Other feature fixtures use deterministic synthetic signals.

Libraries

Package dependencies retain their own licences and notices.

Silero VAD v6.2.3

The embedded silero_vad.onnx release model is distributed under the following licence.

Upstream licence: https://github.com/snakers4/silero-vad/blob/5cd7945676eb32225748052e2e6a0580e4686a08/LICENSE

Model SHA-256: 1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3.

MIT License

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Downloaded Whisper checkpoints

Whisper weights are not embedded. First use downloads the pinned publisher files and converts them locally. The five catalog entries retain the original OpenAI MIT notice, including "Copyright (c) 2022 OpenAI", in the embedded catalog attribution. The pinned Hugging Face cards declare Apache-2.0 for tiny, base, small and large-v3, and MIT for large-v3-turbo. The Apache-2.0 text is included in LICENSE-APACHE. Source repositories, revisions and file digests are recorded in the catalog and model provenance.

Optional speaker checkpoints

The core and tool packages do not include the following publisher checkpoints. RecursiveAI.OneSpeech.Models.CamPlusEnglish includes the pinned English CAM++ publisher files below; without that package they are downloaded and converted on first use of the default diarizer. The other entries are pinned by the development fetch tool. The model cards declare Apache-2.0; the licence text is included in LICENSE-APACHE. These are published pretrained weights, not models trained or fine-tuned by this project. The training-data statements below are the publishers' descriptions, not an independent audit or a licence grant for that data.

SpeechBrain ECAPA-TDNN, speechbrain/spkrec-ecapa-voxceleb revision 0f99f2d0ebe89ac095bcc5903c4dd8f72b367286, embedding_model.ckpt, SHA-256 0575cb64845e6b9a10db9bcb74d5ac32b326b8dc90352671d345e2ee3d0126a2. Source: https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb . Publisher-stated training data: VoxCeleb1 and VoxCeleb2.
CAM++ English — Hui Wang, Siqi Zheng, Yafeng Chen, Luyao Cheng and Qian Chen, published by 3D-Speaker/ModelScope iic under Apache-2.0. iic/speech_campplus_sv_en_voxceleb_16k revision 032b8131a7ad812f87061955ca974c99060c5a03, campplus_voxceleb.bin, SHA-256 5b1a88b6f8d85826fabef804779c3372b42f3af21457fa48bd5c097c0686b2de. Source: https://modelscope.cn/models/iic/speech_campplus_sv_en_voxceleb_16k . Publisher-stated training data: VoxCeleb2, 5,994 speakers.
CAM++ Chinese common, iic/speech_campplus_sv_zh-cn_16k-common revision 013c18884dffb82846cf472dd06dd4a08d97e2cd, campplus_cn_common.bin, SHA-256 3388cf5fd3493c9ac9c69851d8e7a8badcfb4f3dc631020c4961371646d5ada8. Source: https://modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common . Publisher-stated training data: a large Chinese speaker dataset with over 100,000 speakers; the card does not identify the corpus by name.
ERes2NetV2 Chinese common, iic/speech_eres2netv2_sv_zh-cn_16k-common revision 1cf80d41fb3435bd3d8df185b5c423333b2db42a, pretrained_eres2netv2.ckpt, SHA-256 0eb4057106b2573dd7b132cf0c36273ab29afd192c1610f80baa9c556dbb963c. Source: https://modelscope.cn/models/iic/speech_eres2netv2_sv_zh-cn_16k-common . Publisher-stated training data: a large Chinese speaker dataset with approximately 200,000 speakers; the card does not identify the corpus by name.

LibriSpeech evaluation identifiers

eval/manifests/librispeech-speaker-trials.tsv and eval/manifests/librispeech-speaker-parity50.txt contain utterance identifiers derived from LibriSpeech test-clean/test-other (OpenSLR resource 12), by Vassil Panayotov, Guoguo Chen, Daniel Povey and Sanjeev Khudanpur (2015).

Source: https://www.openslr.org/12/ . Dataset licence: Creative Commons Attribution 4.0 International, https://creativecommons.org/licenses/by/4.0/ .

OneSpeech selected and paired these identifiers for evaluation; no original audio or transcript text is included in these files. The selection, pairing method and seed are documented in docs/speakers.md. This attribution does not imply endorsement by the dataset authors.

Reference observations

The tone feature fixtures were produced from deterministic synthetic audio by black-box calls to openai-whisper 20250625 (MIT). They contain only numerical outputs, not model weights, source code or third-party audio. The oracle harness also supports transformers 5.22.2 (Apache-2.0). Neither oracle is a shipped dependency.

LibriSpeech fixture

testdata/1089-134686-0001.flac is from the LibriSpeech ASR corpus (2015), Vassil Panayotov, Guoguo Chen, Daniel Povey and Sanjeev Khudanpur. Test-clean utterance 1089-134686-0001 is distributed unmodified under Creative Commons Attribution 4.0: <https://creativecommons.org/licenses/by/4.0/>. Source: <https://www.openslr.org/12>. The archive path and SHA-256 are recorded in testdata/manifest.json.

ECAPA reference-only auxiliary files

The pinned speechbrain/spkrec-ecapa-voxceleb reference also initializes classifier.ckpt (5,534,328 bytes, SHA-256 fd9e3634fe68bd0a427c95e354c0c677374f62b3f434e45b78599950d860d535) and mean_var_norm_emb.ckpt (1,921 bytes, SHA-256 cd70225b05b37be64fc5a95e24395d804231d43f74b2e1e5a513db7b69b34c33). They are under the same publisher-declared Apache-2.0 licence at revision 0f99f2d0ebe89ac095bcc5903c4dd8f72b367286. These files are used only by the separate evaluation oracle; no classifier or global-embedding normalization weights are included in OneSpeech's converted embedding graph or package.

Qwen3-ASR and Qwen3-ForcedAligner

Qwen Team, Qwen3-ASR and Qwen3-ForcedAligner. The published model cards declare Apache License 2.0. The full licence is provided in LICENSE-APACHE. Model data and converted weights are downloaded separately, not embedded in this package. Pinned checkpoint revisions and file digests are in the model catalog; model provenance and attribution are retained in the returned ModelInfo.

Omnilingual ASR checkpoint and character vocabulary

The optional facebook/omniASR-CTC-300M checkpoint and vocabulary are published by Meta AI / the Omnilingual ASR Team under Apache License 2.0, reproduced in LICENSE-APACHE. Source: https://huggingface.co/facebook/omniASR-CTC-300M/tree/8e35f0cc28fa6099e0c14d56db85ce0423baa691

OneSpeech converts the published tensors to its runtime format without training or fine-tuning. Weights are downloaded separately; they are not embedded in the core package. Retain the licence and these attribution notices when redistributing the model or converted weights.

Moonshine English speech recognition weights

The tiny, base, streaming-tiny, streaming-small and streaming-medium checkpoints from moonshine-ai are MIT-licensed. Their pinned model cards and source links are in the model catalog. The optional .Models.MoonshineTiny, .MoonshineBase and .MoonshineStreamingTiny packages include their respective pinned publisher files (weights, configuration and tokenizer data). These are published pretrained weights, not models trained or fine-tuned by this project. Publisher-stated training data: tiny/base, 200,000 hours of internet audio and open datasets; streaming-tiny, roughly 300,000 hours comprising those sources plus 100,000 hours of internally prepared speech data. These descriptions are not an independent audit or a licence grant for the training data. No Community License checkpoints are included. Publisher licence: https://github.com/moonshine-ai/moonshine/blob/main/LICENSE

MIT License

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

OpenMDW agreement text and Nemotron 3.5 ASR

The OpenMDW-1.1 agreement text is from https://openmdw.ai/license/1-1/; copyright The Linux Foundation and its contributors. The site's contents are licensed under CC-BY-4.0. The agreement is copied without substantive changes in docs/licenses/OpenMDW-1.1.txt and in the model's provenance metadata. Model rights are governed by OpenMDW-1.1, not the site's content licence.

NVIDIA Nemotron 3.5 ASR weights are loaded only by explicit name from the pinned publisher revision and converted in the user's cache. They are not embedded, mirrored, or redistributed by OneSpeech. The full agreement and NVIDIA origin notice are retained in ModelInfo.

NVIDIA Nemotron 3 Diarization (optional publisher download)

nvidia/Nemotron-3-Diarization revision f667ed73aee57d40cc39428eb768b4fd87a0a29e is governed by the OpenMDW License Agreement, version 1.1. No weights or converted graphs are redistributed by this package. Explicit loading fetches the pinned publisher file into the user's cache and constructs runtime artifacts locally. ModelInfo retains NVIDIA's model origin and the full agreement.

Synthetic contextual-tokenizer fixture

parakeet-phrase-tokenizer-reference.json contains outputs we generated from synthetic text using the released tokenizers 0.23.3 public API (Apache-2.0) and NVIDIA's pinned Parakeet tokenizer data. It contains no model weights or copied tokenizer vocabulary. Source: NVIDIA, parakeet-tdt-0.6b-v3, https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 at revision 541d1f99c6b0c3cd0b11a95167540bb8edefd82b. Model licence: CC-BY-4.0, https://creativecommons.org/licenses/by/4.0/. This notice identifies the fixture's provenance; tier-2 models are still fetched only from the publisher and are never bundled. The fixture generator and input texts are ours.