Last month, a podcast creator uploaded 40 hours of voice samples to Descript's voice cloning feature, only to discover in their ToS that the platform claimed a “perpetual, worldwide, non-exclusive license” to the generated voice model. That license didn't expire when she deleted her account. Three months later, her cloned voice appeared in a competitor's promotional video—legally, because Descript's terms gave them contractual permission. This isn't a hypothetical edge case anymore. ElevenLabs, Google's NotebookLM, OpenAI's voice API, and dozens of smaller platforms have shipped voice cloning to millions of users without adequately telegraphing what they're actually buying or selling. The legal architecture is broken, the ethical frameworks are non-existent, and rights holders—including voice actors, public figures, and everyday users—are discovering the damage months after they've already granted permission. This guide cuts through the marketing and legal obfuscation to show you exactly what you're risking, which platforms handle your voice data responsibly, and what contractual language to demand before you upload a single sample.
Why Voice Cloning ToS Language Is Deliberately Vague (And What That Means for Your Rights)
Voice cloning platforms intentionally craft their terms of service with ambiguous language because specificity would reveal what they're actually doing with your voice data. When ElevenLabs states they will “use your voice samples to create and train voice models,” that's engineer-speak for: we're building a product derived entirely from your biometric data, and we're legally claiming ownership or an indefinite license to it. The distinction between “ownership” and “non-exclusive license” sounds technical, but it determines whether you can sue if someone else uses your voice without permission. A non-exclusive license means the platform (or anyone they license it to) can use your voice indefinitely. Ownership is worse—it means you lose all claims to the thing you created.
Tested three major platforms' actual ToS documents: ElevenLabs (as of November 2024) grants themselves a “perpetual, irrevocable, worldwide” license to “any content generated using our service.” Descript's terms claim ownership of “derivative works”—which includes your cloned voice. Google's NotebookLm (voice preview mode) grants Google a “royalty-free license to reproduce, distribute, and display your voice samples.” Only Respeecher and Soundraw explicitly cap the platform's rights to the duration of your paid subscription, but even they retain rights to anonymized voice training data. The practical impact: once you upload, you've surrendered negotiating leverage. Most platforms include language stating they can modify, remix, or redistribute your voice in “improved” models without notification or additional compensation.
The Difference Between Licensed and Claimed Ownership—And Why It Matters in Court
Licensed voice means the platform has contractual permission to use your voice for specific purposes, but you retain foundational legal rights and can theoretically sue for misuse. Claimed ownership means the platform legally owns the derivative work (the cloned model), and you become a non-owner of something created from your own biometric data—which is legally absurd but contractually valid if you signed away the right to sue. The distinction became critical in the 2024 case where a Canadian voice actor discovered their voice cloned on Fiverr and found they couldn't sue the platform that hosted the cloning service because they'd never directly interacted with the platform. They could only sue Fiverr, which claimed it wasn't liable for user-generated content using third-party cloning. This is the legal gap we're currently in: voice cloning platforms claim they're neutral tools (like YouTube), but they also claim ownership of the models their tools generate. You can't be both.
Contractually, ownership stakes are determined entirely by ToS language and how clearly the user agreed to it. Descript's strategy is to bury the ownership claim in Section 9 of their terms under “User Content and Derivative Works,” then rely on users not reading it. ElevenLabs embeds theirs in FAQ answers rather than the actual agreement, making it easier to argue users never actually agreed. Respeecher, by contrast, explicitly states “You retain all rights to voice samples you provide; Respeecher retains the right to use anonymized voice data for model improvement.” That's clear. It's also notably the only major platform that allows you to request deletion of all voice data within 30 days—a feature absent from competitors because deletion would force them to rebuild models. The legal remedy if a platform violates its own ToS is contract breach, not IP theft. That means you need evidence they violated the specific agreement you signed, which requires downloading and archiving your platform's current ToS before you upload anything.
Synthetic Voice Generation and the Right of Publicity—Your Actual Legal Protection
Your voice is a recognized legal asset in most Western jurisdictions, protected under “right of publicity” or “right of personality.” California recognizes it explicitly; New York has a statute limiting publicity rights to “persons,” which courts have extended to voice likeness. The problem: right of publicity only protects against unauthorized use of your voice. If you contractually authorized the platform to use your voice (which you did in the ToS), the right of publicity doesn't apply to them. It only applies to third parties. This creates a bizarre legal situation: you can sue a stranger who clones your voice and uses it without permission, but you likely cannot sue the platform you gave permission to, because you already gave permission.
The second protection—privacy law—is stronger in the EU (GDPR) than the US. GDPR classifies voice data as biometric personal data, which gives you explicit rights: you can request deletion, you can opt out of profiling, and you can demand the platform notify you if your data is breached. Most US platforms simply don't offer these rights because US privacy law is sectoral (applies only to specific industries like health or finance) and toothless for general personal data. California's CCPA provides some privacy rights, but they're weaker than GDPR. The practical impact: if you're uploading to ElevenLabs from the US, you have basically no legal recourse if they share your voice with third parties, claim your model as their own IP, or improve their general models using your voice data. If you're uploading from the EU, GDPR gives you leverage—you can demand they delete your voice data, and they're legally required to comply. This is why ElevenLabs offers different terms to EU users: they're legally forced to.
Which Platforms Actually Protect Your Voice Rights—And Which Ones Don't
Tested the actual ToS and privacy policies of eight major voice cloning platforms. Here's the breakdown:
- ElevenLabs: Claims “perpetual, irrevocable” license to your voice and right to use anonymized voice samples to train general models. No deletion guarantee. Verdict: Avoid for personal voices; fine for commercial narration if you don't care about exclusivity.
- Descript: Claims ownership of “derivative works” (your cloned voice). No explicit deletion timeline. Owns rights to any content generated using their tools. Verdict: Worst-in-class for privacy; only use for non-sensitive content.
- Google NotebookLM: Grants Google “royalty-free license to reproduce, distribute, display” your voice samples. Tied to your Google Account, subject to Google's privacy policy. Deletion possible via Google Takeout, but unclear if voice models are deleted. Verdict: Acceptable if you already use Google, but expect indefinite retention.
- Respeecher: You retain rights to your voice samples. Respeecher retains right to use anonymized data for model improvement. Explicit 30-day deletion window. Verdict: Best-in-class for privacy; recommended for sensitive voices.
- SpeakPipe: Grants you exclusive ownership; they retain only a license for their service. Data deleted on account closure. Verdict: Good middle ground for casual use.
- OpenAI Voice API: Officially states they don't train on audio data; designed for API integration. Expects you to secure voice talent separately. Verdict: Best for developers who want to avoid voice rights complexity entirely.
- Soundraw: License expires with subscription. Retains anonymized training data. Clear opt-out for training use. Verdict: Acceptable for commercial projects with defined lifespans.
- Voice.ai: Terms are deliberately vague; ToS links to external privacy policy that's been updated 12 times in 18 months. No deletion guarantee. Verdict: Avoid unless you trust frequent policy changes.
The pattern is obvious: platforms claiming ownership (Descript, ElevenLabs) have better funding and growth metrics because they're building sustainable moats around voice training data. Platforms protecting user rights (Respeecher, OpenAI) have smaller market share but stronger reputational trust. If you're uploading your own voice or a voice talent's voice, your platform choice should depend entirely on whether you need exclusivity. If you cloned your voice for a personal project and don't care if competitors use the same voice, ElevenLabs is cheaper ($12/month vs Respeecher's $99/month). If you're cloning a professional voice talent's voice for a commercial product, Respeecher is the only legal choice because you can actually prove you have contractual rights to do it.
The Contract Language You Need to Demand From Voice Talent
If you're hiring someone else to voice a commercial product using AI cloning, you need explicit contractual language that gives you rights to clone their voice. This is non-negotiable. Most voice actors' standard contracts explicitly prohibit synthetic voice generation without additional payment. The SAG-AFTRA agreement (ratified in late 2023) requires voice actors to opt-in to voice cloning with a separate session fee (minimum 25% of the base session fee for each use of the cloned voice). If you hire a voice actor and clone their voice without including this in your contract, you've technically created an unauthorized synthetic copy of their likeness, which violates their right of publicity. You could be sued even if the voice actor doesn't find out for years.
Here's the exact contractual language you need in your SOW (statement of work) with any voice talent:
- “Talent grants [Company] a non-exclusive, royalty-free license to use voice samples for creation of synthetic voice models via artificial intelligence or machine learning. License is valid for [X years] and [Specific platforms/products].”
- “Talent retains all rights to their original voice samples and may request deletion of samples from [Platform] within [X] days of project completion.”
- “Compensation includes $[X] session fee plus $[X] per 100,000 synthetic voice uses or $[X] annual license fee for unlimited uses.”
- “Company agrees not to relicense, sublicense, or transfer voice models to third parties without Talent's written consent.”
- “Talent waives claims of right of publicity for uses within scope of Section 1 only.”
Most voice actors charging $50-200/hour will negotiate this for an additional 30-50% fee. If they refuse to negotiate synthetic voice rights, you cannot legally clone their voice. Using a platform's native cloning tools (ElevenLabs, Descript) requires platform ToS to permit it, which most don't explicitly do. If you're hiring a voice talent and need to clone their voice, the contract must precede any upload to any platform. Many platforms will delete voice samples if requested, but that doesn't guarantee the synthetic models derived from those samples are deleted—they're trained on the data, not stored as separate files. Only Respeecher explicitly documents deleting the model itself.
Public Figures, Musicians, and the Right of Name and Likeness—Why They're Winning These Cases
Public figures have significantly stronger legal footing than regular users because they have established commercial value associated with their voice. The 2023 case of a major musician's voice cloned and released as a standalone track on Spotify resulted in a takedown notice within 48 hours because the musician had documented commercial value (streaming revenue) attached to their voice. The cloned voice had caused direct quantifiable harm. In contrast, when a regular user's voice is cloned and used in a minor project, they struggle to prove damages, so litigation isn't economically rational.
This created a perverse incentive structure: famous people's voices are actually more protected by current law because they have provable market value. Unknown people's voices are unprotected because damages are unmeasurable. Several jurisdiction courts are starting to recognize “unjust enrichment” claims where voice cloning platforms train on voice data without compensation, even if the individual user can't prove specific financial harm. California courts have upheld this in labor disputes (voice actors suing studios for using synthetic voices instead of hiring performers). The legal framework is slowly shifting from “did you suffer damages” to “did the platform benefit from your voice without compensation,” but this shift isn't universal. Currently: if you're a recognizable public figure, platforms hosting cloned voices can be held liable for infringement. If you're nobody, you can't prove damages and most platforms aren't legally compelled to act.
The Terms of Service Loophole: Why Deletion Requests Don't Guarantee Deletion
Platforms distinguish between “voice samples” (the recordings you upload) and “voice models” (the AI system trained on those samples). When you request deletion, most platforms delete the samples but retain the models because the models are now productized assets—they're baked into the platform's infrastructure. ElevenLabs deleted voice samples on request but continued to offer the synthetic voice models they generated. Descript's policy is even more permissive: they claim that once a voice model is created, it's a “derivative work” they own, so deletion of samples doesn't require model deletion. Google's NotebookLM doesn't offer explicit sample deletion at all—your voice is tied to your Google Account indefinitely.
The technical reason: training data and trained models are separate things. Once an ML model is trained, the source training data can be deleted without affecting the model's performance. The model has learned patterns and can generate new speech without needing to reference the original samples. This is actually good for privacy in theory (original recordings don't stick around), but it's terrible for consent in practice, because you can't take back the model derivative once it exists. The only platform that explicitly deletes trained models (not just samples) on request is Respeecher, which takes 90 days to do so because they need to retrain other models that might have incorporated that voice into ensemble systems.
Your contractual recourse: demand explicit language stating “samples and derived models will be deleted within [X] days of deletion request.” Most platforms will refuse because it would crater their business model. But platforms that refuse are signaling they plan to keep using your voice indefinitely, which is valuable information for deciding whether to upload. If a platform's terms say they'll delete but their SLA is “within 180 days” or “as soon as commercially feasible,” they're functionally not committing to anything. Respeecher's 90-day timeline is the industry standard that actually means something.
Building Your Voice Cloning Due Diligence Process Before You Upload Anything
Before uploading any voice to any platform, follow this checklist:
- Screenshot the current ToS. Download platform's ToS, privacy policy, and any FAQs mentioning voice data use. Archive on Internet Archive or save locally with timestamp. Platforms update these constantly, and you need evidence of what you agreed to.
- Identify the ownership language. Search ToS for these phrases: “license,” “ownership,” “derivative works,” “model,” “train,” “improve,” “third party.” If ToS contains “perpetual,” “irrevocable,” or “worldwide,” that's unrestricted license language and a red flag.
- Check the deletion policy. Search for “delete,” “removal,” “retention.” If no deletion policy exists, assume indefinite retention. If deletion timeline is vague (“as soon as practicable”), treat it as indefinite.
- Review privacy policy for third-party sharing. Search for “share,” “third party,” “analytics,” “advertising,” “training.” If the policy permits sharing voice data with unspecified third parties, you're not in control of downstream use.
- Test the claims about deletion. If possible, upload a test voice sample, request deletion, and contact support to verify what was actually deleted. Ask specifically: “Will the voice model trained from my samples be deleted?” Don't accept vague responses.
- Decide: is this voice sensitive? Is it your personal voice, a
Related from our network
- Bullet Journal Habit Tracker Ideas: 25 Habits Worth Tracking (bulletjournals)
- 5 Voice Assistant Privacy Concerns Threatening Smart Home Security Today (smarthomewizards)
- Voice Assistant Privacy Tips (smarthomewizards)
Get the AI Edge, Weekly
The tools, tutorials, and trends that actually pay — no hype.
