Appendix C. Toolkits, models, data and their licences¶
Licences change, and a model's weights can carry a different licence from the code that runs them. Each entry
below gives what the named source said when this list was made, on 2026-10-29: the repository's licence as
GitHub reports it, the license field of a Hugging Face model or data card, or the line quoted in the
chapter given. Read the licence itself before you ship anything.
Toolkits and libraries (code)¶
| Project | Licence | Used in |
|---|---|---|
| k2-fsa/sherpa-onnx | Apache-2.0 | CPU-track recognition, chapters 3, 7, 9, 15 |
| k2-fsa/icefall | Apache-2.0 | recipes, export, fine-tuning; chapters 3, 5, 15, Appendix E |
| csukuangfj/kaldi-native-fbank | Apache-2.0 | features, chapter 2 |
| lhotse-speech/lhotse | Apache-2.0 | features, chapter 2 |
| NVIDIA/NeMo | Apache-2.0 | GPU-track recognition, chapters 5, 9, 15 |
| wenet-e2e/wenet | Apache-2.0 | chapters 2, 5, 9 |
| openai/whisper | MIT | chapters 2, 8, 9 |
| ufal/whisper_streaming, ufal/SimulStreaming | MIT | chapter 8 |
| jitsi/jiwer | Apache-2.0 | scoring, chapter 9 |
| snakers4/silero-vad | MIT (the README badge's alt-text says "CC BY-NC 4.0"; the LICENSE file and the badge image say MIT) | chapters 7, 18 |
| pipecat-ai/pipecat | BSD-2-Clause | chapters 7, 18 |
| livekit/agents | Apache-2.0 | chapters 7, 18 |
| hexgrad/kokoro, hexgrad/misaki | Apache-2.0 | CPU-track synthesis, chapters 10, 12 |
| espeak-ng | GPL-3.0 | Kokoro's G2P fallback and non-English G2P, chapter 10 |
| FunAudioLLM/CosyVoice | Apache-2.0 | chapters 10 to 14 |
| QwenLM/Qwen3-TTS | Apache-2.0 | chapters 11, 12 |
| vllm-project/vllm-omni | Apache-2.0 | GPU-track synthesis serving, chapters 12, 16 |
| kyutai-labs/moshi | MIT (Python), Apache-2.0 (Rust), per its README | chapters 11, 12 |
| kyutai-labs/delayed-streams-modeling | MIT (Python), Apache-2.0 (Rust), per its README | chapters 8, 12 |
| facebookresearch/encodec | MIT (code) | chapter 11 |
| descriptinc/descript-audio-codec | MIT, code and weights ("Weights are released as part of this repo under MIT license") | chapter 11 |
| hubertsiuzdak/snac | MIT | chapter 11 |
| gemelo-ai/vocos | MIT | chapter 13 |
| jik876/hifi-gan, NVIDIA/BigVGAN | MIT | chapter 13 |
| SWivid/F5-TTS | MIT (code) | chapters 10, 13 |
| coqui-ai/TTS | MPL-2.0 (code) | chapters 12, 13 |
| canopyai/Orpheus-TTS | Apache-2.0 (code) | chapters 11, 16 |
| SesameAILabs/csm | Apache-2.0 | chapter 12 |
| FireRedTeam/FireRedTTS | MPL-2.0 | chapter 12 |
| BytedanceSpeech/seed-tts-eval | no licence file | chapter 14 |
Model weights¶
| Model (Hugging Face) | Revision | Card licence | Used in |
|---|---|---|---|
| csukuangfj/sherpa-onnx-streaming-zipformer-en-2023-06-26 | 672fbf1 | apache-2.0 | CPU track, English |
| csukuangfj/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20 | 98590b7 | apache-2.0 | CPU track, serving |
| k2-fsa/sherpa-onnx-streaming-zipformer-multi-zh-hans-2023-12-12 | ac54a23 | apache-2.0; trained on data sets that include WenetSpeech, AISHELL-2 and KeSpeech, whose terms restrict them to research or non-commercial use (see below) | CPU track, Mandarin |
| nvidia/stt_en_fastconformer_hybrid_large_streaming_multi | ae98143 | cc-by-4.0 | GPU track, English |
| nvidia/nemotron-3.5-asr-streaming-0.6b | ea30d66 | OpenMDW-1.1 | GPU track |
| openai/whisper-large-v3 | 06f233f | apache-2.0 | scoring synthesized English, chapter 14 |
| funasr/paraformer-zh | d7811ee | apache-2.0 | scoring synthesized Mandarin, chapter 14 |
| pipecat-ai/smart-turn-v3 | f766f81 | bsd-2-clause | turn detection, chapters 7, 18 |
| livekit/turn-detector | fba34c3 | "LiveKit Model License"; its LICENSE file allows use "only ... together with the LiveKit Agents framework" | chapter 7 |
| hexgrad/Kokoro-82M, hexgrad/Kokoro-82M-v1.1-zh | f3ff357, 01e7505 | apache-2.0 | CPU-track synthesis |
Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice |
0c0e305 | apache-2.0 | GPU-track synthesis |
| Qwen/Qwen3-1.7B | 70d244c | apache-2.0 | the language model of chapter 18 |
| FunAudioLLM/CosyVoice2-0.5B, FunAudioLLM/Fun-CosyVoice3-0.5B-2512 | eec1ae6, 29e01c4 | apache-2.0 | chapter 12 |
| kyutai/mimi | 89091b3 | cc-by-4.0 | chapters 11, 13 |
| kyutai/tts-1.6b-en_fr, kyutai/stt-1b-en_fr, kyutai/stt-2.6b-en | f654396, 1c34c6b, a07aec5 | CC-BY 4.0 (the cards: "Model weights are licensed under CC-BY 4.0") | chapters 8, 12 |
| mistralai/Voxtral-Mini-4B-Realtime-2602 | 2769294 | apache-2.0 | chapter 8 |
| SWivid/F5-TTS | 84e5a41 | cc-by-nc-4.0 (non-commercial, "due to the training data Emilia", per the GitHub README) | chapter 10 |
| coqui/XTTS-v2 | 6c2b0d7 | Coqui Public Model License | chapter 12 |
| canopylabs/orpheus-3b-0.1-ft | 4206a56 | apache-2.0; its base model is meta-llama/Llama-3.2-3B-Instruct, under Meta's own licence | chapter 16 |
| sesame/csm-1b | c92a71e | apache-2.0 | chapter 12 |
| descript/dac_24khz | 6ba020b | not stated on the card; the code repository says the weights are MIT | chapter 11 |
| facebook/encodec_24khz | c1dbe2a | not stated on the card | chapter 11 |
Data sets¶
| Data set | Licence | Note |
|---|---|---|
| LibriSpeech | CC BY 4.0 (openslr.org/12) | English test sets of chapters 1 to 9, 15 |
| LibriTTS-R | cc-by-4.0 (Hugging Face card) | English clips of chapters 11, 13, 14 |
| AISHELL-1 | Apache License 2.0 (openslr.org/33) | Mandarin test sets throughout |
| FLEURS | cc-by-4.0 | multilingual evaluation |
| WenetSpeech | non-commercial, per its official site ("available to download for non-commercial purposes under a Creative Commons Attribution 4.0 International License"), although openslr.org/121 lists plain CC BY 4.0 | not used directly; in the training data of the CPU-track Mandarin model |
| GigaSpeech | non-commercial research and education, per its access terms, although its Hugging Face metadata says apache-2.0 | not used here |
| TED-LIUM 3 | CC BY-NC-ND 3.0 | not used here |
Where two sources disagree, the stricter terms are the ones the data is released under: the metadata tag describes the repository, the access terms describe the audio.
A model's weights can be released under a permissive licence while its training data is not. The CPU track's Mandarin model is one: its icefall recipe (multi_zh-hans) trains on about 14,000 hours that include WenetSpeech, AISHELL-2 and KeSpeech. The book uses it because it is a small streaming Mandarin model for sherpa-onnx that reached 4.55% CER on the book's AISHELL-1 test subset (chapter 9); before you build a commercial product on it, read the terms of those data sets and decide whether they bind the model.