Licensed data, consented voice talent: 6 questions for your TTS vendor

Ask your TTS vendor six things before you ship a synthetic voice: what the training data was, whether the voices are consented, how cloning is gated, whether the audio is watermarked, what the commercial terms are, and which jobs the voice was tuned for. Three vendors made public statements on this in the last two weeks, and they differ.

What each vendor says on its own pages

The statements below come from each vendor's own page, read on 2026-10-07.

The six questions

Use these as a checklist in a vendor email.

• What data trained the voice, and was the voice talent paid and consenting?
• Can a customer clone a voice, and what proof of consent does the vendor require?
• Is the output watermarked, and does the mark survive trimming and re-encoding?
• What are the commercial terms for ads, and is the license for the weights or the API?
• Was the voice tuned for your use, such as narration, or for another, such as support calls?
• Which terms apply if you modify the voice or edit the output?

Why the last two matter most

Decagon describes Chord as post-trained on real-world customer-experience conversations, which points to support calls. Microsoft says MAI-Voice-2.1 prioritizes expression over latency, and its Flash variant is aimed at agents. A vendor's tuning tells you where it is likely to shine and where it will not.

The license matters too: Mistral lists Voxtral TTS weights as CC BY-NC 4.0, which is non-commercial, while the paid API is a separate route.

Where Sume fits

Sume's TTS surface takes a voice selector from the workspace's voice library or an avatar, and it records what it used. A completed text_to_speech job stores the model id, voice, language, output format and the synthesis settings, so you can show which voice made which file. For consent and licensing questions about a specific voice, ask Sume support before you rely on it in a campaign; this post only reports what the job record shows.