What the licence of a local speech model actually lets you do

“Local model” sounds like one legal category. It is three. The engines that a Mac transcription app ships come from OpenAI, NVIDIA and Alibaba’s FunAudioLLM, and their licences differ enough that the answer to “may I do this” changes depending on which one is selected in the menu.

I am not a lawyer and this is not legal advice. Below is what the three licences say, quoted, with links to read them yourself.

Three models, three licences

Model	Licence	Commercial use	Attribution
Whisper (large-v3-turbo, small)	MIT	Yes	Keep the copyright notice with the software
Parakeet TDT (v3, CTC 110M)	CC BY 4.0	Yes, stated on the model card	Credit NVIDIA
SenseVoiceSmall	FunASR Model License	Not addressed either way	Attribute source and author, keep model names

Whisper is the permissive end. The Whisper license is plain MIT: it grants “permission… free of charge, to any person obtaining a copy of this software and associated documentation” to use it, with the notice preserved. There is nothing to negotiate.

Parakeet asks for credit. The Parakeet model license is CC BY 4.0: NVIDIA’s model card states that “use of this model is governed by the CC-BY-4.0 license” and that the model is ready for commercial and non-commercial use. CC BY is permissive in substance and specific about attribution.

SenseVoice is the one to actually read. The SenseVoice license — formally the FunASR Model License — is not a standard open-source licence, and it has a clause I have not seen in any other model licence:

you must attribute the source and author information and retain relevant model names

followed by a prohibition on “unjustified denigration, malicious smearing, or baseless insults”, with the licence forfeited automatically on violation. Whether that is enforceable anywhere is a question for someone with a bar admission. That it is unusual is not.

The column that is empty in every row

Here is the finding that took the reading to reach. A speech model license, any of these three, says nothing at all about the transcript.

Not “the output belongs to you”, not “you may not sell it”, not a word. MIT governs software. CC BY governs the model as a licensed work. The FunASR licence governs the weights and derivatives. Ask any of them who owns the text that comes out and they are silent, because they were written to answer a different question.

This matters because the question people actually have is about the transcript. A freelancer asks whether they may charge for subtitles made with Whisper. A researcher asks whether interview transcripts made with Parakeet can go in a published paper. Both go looking in the licence, and the licence is not where the answer lives.

Where the answer does live

Two other questions decide it, and neither has anything to do with the model.

Who holds rights in the recording. This is the one that bites. A transcript of a film is a derivative of that film’s dialogue, and no model licence changes that. A transcript of your own meeting is yours to do as you like with, subject to whatever the other participants agreed to. Swapping Whisper for Parakeet does not move that line by a millimetre — it is the same line whether you transcribe by machine or type it out by hand.

Whether machine output is copyrightable at all. The US Copyright Office addressed this directly in its March 2023 registration guidance (88 FR 16190). Its position, quoting its own Compendium, is that “to qualify as a work of ‘authorship’ a work must be created by a human being”, and that it “will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author”.

Read that against a transcript and the picture is odd but coherent. The transcription itself — a machine converting speech to text with no creative choices — is not the sort of thing the Office registers. The speech it transcribed may well be protected, and its author is the person who spoke. Which returns you to the first question and away from the model entirely.

What this means in practice

Transcribing your own recordings: nothing in any of the three licences constrains you. Attribution clauses attach to the model, not to text it produced, and they become concrete when you redistribute the model or ship it inside something.

Selling transcription as a service: all three permit commercial use of the software — Whisper explicitly, Parakeet explicitly on the model card, SenseVoice by not prohibiting it. Your obligations are attribution ones, satisfied by naming the engines in your documentation.

Building the model into a product: this is the case the licences were written for, and the one to get right. Credit NVIDIA for Parakeet, keep the MIT notice with Whisper, attribute source and author and retain model names for SenseVoice. It is a page in your docs, not a lawyer’s afternoon.

Transcribing material you did not record: the licences will not save you and never claimed they would. That is a rights question about the source, and it is the same question it was before speech models existed.

If you want to see how this is done concretely: Kekoso lists every model it can run in its About window with the vendor and the licence name, each linking to the licence text, so a person who wants to know which of these three they are running does not have to go looking for it.

---

Which of these engines is worth choosing on the merits rather than the paperwork is in Whisper vs Parakeet vs SenseVoice. If your interest in the legal side is about recording people in the first place, the statutes on call recording are quoted in the same way as the licences here.

Questions people ask

Can I use Whisper commercially?

Yes. Whisper is released under the MIT licence, which grants permission to use, copy, modify and distribute the software free of charge, commercial use included, provided the copyright notice travels with it. What the licence does not do is say anything about transcripts — it governs the software, not what comes out of it.

What licence is NVIDIA Parakeet under?

CC BY 4.0. NVIDIA's model card states that use of the model is governed by that licence and that the model is ready for commercial and non-commercial use. CC BY carries an attribution requirement, which matters if you redistribute the model or build it into a product, and is largely moot if you are simply transcribing your own files.

Does SenseVoice have the same licence as Whisper?

No. SenseVoiceSmall ships under the FunASR Model License, which is not one of the standard open-source licences. It requires that you attribute the source and author and retain the model names, and it contains an unusual clause: unjustified denigration or malicious smearing of the software forfeits the licence automatically.

Do I own the transcript a local model produces?

The model licences do not answer that, because none of the three mentions outputs at all. The questions that actually decide it are different ones: who holds rights in the recording you transcribed, and whether a machine-produced text is copyrightable in the first place. The US Copyright Office will not register works produced by a machine without creative input from a human author.

Can I sell transcripts made with Whisper?

Nothing in the three licences prohibits it, and Whisper's MIT terms and Parakeet's commercial-use statement are explicit that the models may be used commercially. The constraint that bites in practice is the source material — subtitles for a film you do not have rights to remain a rights problem no matter which model produced them.

Do I have to credit the model in my transcripts?

Attribution under CC BY and the FunASR licence attaches to the model, not to every line of text it produces, and it becomes concrete when you redistribute the model or ship it inside something. Naming the engine in your product's documentation is the ordinary way to satisfy it. Crediting NVIDIA at the bottom of a meeting transcript is not what those clauses are asking for.