Which local speech-to-text models can you use commercially?

All 26 reviewed local speech-to-text models can be used commercially under their listed licenses, but the obligations differ. 22 use MIT or Apache-2.0 (keep the license and notices); Parakeet TDT 0.6B v3 and Parakeet Redux use CC BY 4.0 (credit required); Orukeet uses CC BY-SA 4.0 on its weights (credit plus ShareAlike for adaptations); and Nemotron 3.5 ASR Streaming 0.6B uses OpenMDW-1.1, an open model license that requires keeping the license and notices when you redistribute.

This is a practical comparison of published license terms, not legal advice. Licenses are reviewed against the linked primary sources (latest model review November 3, 2026); always read the current text before you ship.

What each license allows and requires

Grouped by license. "Allows" and "main obligations" summarise the license text itself; follow the link for the full terms.

License of every reviewed model

4 models carry more than notice-keeping obligations: Parakeet TDT 0.6B v3, Orukeet, Nemotron 3.5 ASR Streaming 0.6B and Parakeet Redux. Filter the table by language or streaming if you only need a subset.

No reviewed model matches these filters. Widen the license class or pick another language.

Weights, code and runtimes can have different licenses

• Weights versus repository code. A model repository can license its code and its weights separately. Orukeet keeps its code under MIT but publishes a dedicated LICENSE-WEIGHTS ↗ file: the CC BY-SA 4.0 ShareAlike condition applies to the weights. OpenAI, by contrast, states ↗ that both Whisper's code and its weights are MIT.
• Same project, different checkpoints. Moonshine's license note ↗ says its models are MIT by default, except the legacy non-streaming models for languages other than English, which use a non-commercial community license. The two English Moonshine models reviewed here are MIT.
• Inference runtimes are separate software. faster-whisper, whisper.cpp, NVIDIA NeMo, vLLM or Transformers each have their own license. Check the runtime you deploy as well as the model.
• Fine-tunes inherit the base model's terms. A fine-tuned checkpoint of a CC BY or CC BY-SA model is adapted material: credit the original, and for ShareAlike release your adaptation under the same or a compatible license if you share it.
• No trademark rights. Apache-2.0 and the Creative Commons 4.0 licenses state that they do not grant trademark rights. Name the model accurately, but do not brand your product as if it were the vendor's.

Shipping checklist

1. Record the exact model repository, revision and the license file you relied on, with the date you checked it.
2. Put the license text and any copyright or NOTICE content into your product's third-party notices.
3. For CC BY / CC BY-SA models, add a visible credit line that names the model and creator, links the license and says whether you modified the weights.
4. If you fine-tune a ShareAlike model and distribute the result, publish it under CC BY-SA 4.0 or a compatible license.
5. Check the license of the runtime, any quantized conversion you download and any diarization or alignment add-on separately.
6. Re-check before each major release; vendors occasionally change or clarify terms.

Common questions

Can I use Whisper commercially?

Yes. OpenAI publishes Whisper under the MIT License, which allows commercial use, modification and redistribution as long as the copyright and permission notice are kept with copies of the software. All five reviewed Whisper sizes and Distil-Whisper large-v3 carry MIT.

Is Parakeet free for commercial use?

Parakeet TDT 0.6B v3 is released under CC BY 4.0. Commercial use and adaptation are allowed, but you must give appropriate credit, link the license and indicate changes. The 1.58-bit Parakeet Redux keeps the same CC BY 4.0 license, while the Parakeet-derived Orukeet weights add a ShareAlike condition (CC BY-SA 4.0).

Which local speech-to-text models have the lightest license obligations?

22 of the 26 reviewed models use MIT or Apache-2.0: Whisper (tiny, base, small, medium, turbo), Qwen3-ASR 0.6B, Qwen3-ASR 1.7B, Distil-Whisper large-v3, Moonshine Tiny Streaming (English), Moonshine Medium Streaming (English), Omnilingual ASR CTC 300M, Cohere Transcribe 03-2026, VibeVoice-ASR 7B, VibeVoice-ASR Streaming 1.5B, VibeVoice-ASR-BitNet, Fun-ASR-MLT-Nano-2512, Voxtral Mini 4B Realtime 2602, Granite Speech 5.0 2B, Granite Speech 5.0 470M TurboCTC, VibeVoice-ASR-Streaming-7B, Whistle and Jais-ASR. Both mainly require keeping license and copyright notices; Apache-2.0 also asks you to keep NOTICE files, mark modified files and includes an explicit patent license.

Do I own the transcripts the model produces?

MIT, Apache-2.0 and the Creative Commons licenses govern the model material itself (code and weights). None of their texts claims rights in transcripts you generate. Rights in the underlying audio, and privacy obligations for recorded people, are a separate question that the model license does not answer.

Does a self-hosted transcription service count as distribution?

The notice and ShareAlike duties in these licenses are triggered mainly when you share copies or adapted versions of the licensed material. Running a model on your own server is usually not the same as sharing its weights — but shipping weights inside a desktop app, container image or device is. Read the exact license text for your case; this guide is not legal advice.

Primary sources

License classes come from the reviewed model database. Obligation summaries were written on November 2, 2026 from the license texts linked above. See also the source directory and each model profile.