License terms for speech and voice recordings: speaker consent scope and cloning limits

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why voice data needs a different grant than text

Voice recordings carry two things a transcript does not: a reproducible identity and a potential biometric. The same WAV file can train a recognizer, a speaker-embedding model or a synthetic voice, and those uses carry very different legal and reputational exposure. A generic "machine learning purposes" grant, copied from a text-data template, collapses all of them together.

Statutes treat the voiceprint, not the recording, as the regulated object. Illinois BIPA lists a voiceprint as a biometric identifier, requires a written release before collection, a published retention schedule, and bars selling or otherwise profiting from biometric data [1]. Texas requires notice and consent before capturing a voiceprint for a commercial purpose [2]. Washington's definition covers data generated by automated measurement used to identify a person, while excluding audio recordings themselves [3]. As of October 2026, class actions filed in May 2026 allege that voiceprints were used to train AI in violation of BIPA, and none had been decided on the merits [4].

Your general rights-grant language belongs on the AI training rights grant clause page. This page covers the voice-specific limits that sit on top of it.

Map each permitted use to what speakers consented to

The permitted-use clause should be no wider than the narrowest consent in the corpus, or the corpus should be split by consent tier. Ask the licensor for the consent instrument behind every source: an IVR disclosure, a studio release, a contributor agreement, or an employee policy. Then map each one to the uses below.

Illustrative example: invented to show structure; it does not describe an available dataset.

The voice talent consent and release terms page covers what a studio release must contain. For call audio, start with call-recording consent checks, because all-party consent states such as California require every party's consent to record a confidential communication [9].

Write the voice-cloning limit as a negative covenant

A no-cloning clause works best as an explicit prohibition with a defined test, not an omission from the grant. Omission leaves room to argue that cloning is "training" under a broad definition. Define "Voice Clone" as any model, adapter, speaker embedding or prompt that can generate speech a reasonable listener would attribute to a specific licensed speaker.

Cover the adjacent routes too: zero-shot TTS prompting with licensed clips, fine-tuning a LoRA on one speaker, and exporting x-vectors or ECAPA embeddings as a reusable "voice." Right-of-publicity law is moving here. Tennessee's ELVIS Act of 2024 extends protection to an individual's voice, including AI simulations [10]. Downstream, the FCC treats AI-generated voices in outbound calls as subject to TCPA artificial-voice rules [11], so a cloned voice that reaches a dialer creates a second exposure.

Researchers have documented that voice contributors face long-tail consent, credit, compensation and reputation risks when recordings are reused for generation [7]. Many TTS systems also train on internet-gathered speech where consent and licensing are hard to verify at scale [8]. Those are the failure modes your covenant should rule out.

Ban speaker identification and voiceprint extraction

Unless the licensor can produce written biometric releases, the license should prohibit identifying speakers, building enrollment templates and linking audio to external identity. That prohibition reduces exposure under BIPA's written-release and profit restrictions [1] and Texas's capture rule [2]. Write it as conduct-based, because a stored speaker embedding can function as a voiceprint even if no one calls it one.

Practical terms to include:

• No speaker verification, identification or re-identification models trained on licensed audio.
• No retention of per-speaker embeddings beyond a training run, and no export outside the training environment.
• No joining of audio to account numbers, ANI/caller ID, CRM records or employee directories.
• Speaker labels in RTTM or diarization files are pseudonymous and rotate per release.
• Anonymization claims are measured, not asserted. VoicePrivacy-style evaluation reports attacker equal error rate alongside ASR word error rate on anonymized speech [5]. See speaker anonymization metrics.

GDPR is stricter about the endpoint: data is outside its scope only when a person is no longer identifiable by means reasonably likely to be used [13]. A raw voice rarely meets that bar, so EU-linked audio usually remains personal data even after transcript redaction.

Separate retention rules for audio, transcripts and derived features

Raw audio, transcripts and derived features carry different risk, so the license should set retention for each. Raw audio holds the voice and any spoken PII that redaction missed. Redacted transcripts hold far less. Derived features (MFCCs, log-mel spectrograms, codec tokens) sit in between, since some can be inverted back to intelligible speech.

A workable structure: raw audio deleted or returned at term end or on speaker revocation; transcripts retained for the license term; features governed like audio unless shown to be non-invertible. BIPA's retention-schedule duty applies to biometric data you hold [1], so any embedding store needs its own destruction date. The deletion and return clause and model retention after termination pages cover the trained-model question; for voice, ask specifically whether a TTS checkpoint that can still render a revoked speaker's voice must be retrained.

Diligence evidence to request before signing

Ask the licensor to deliver consent evidence at the clip level, not as a policy summary. An audit of text datasets found popular hosts omitted licenses for over 70% of them, with license error rates above 50% [15]; audio collections deserve the same skepticism. Even a well-known public-domain corpus like Common Voice was collected for speech recognition [6], which tells you nothing about its suitability for cloning a contributor.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "clip_id": "c-000481",
  "source_type": "inbound_support_call",
  "recording_disclosure": "ivr_v3_ai_training",
  "consent_jurisdiction": "US-CA",
  "all_party_consent": true,
  "permitted_uses": ["asr", "diarization", "voice_agent"],
  "synthesis_consent": false,
  "biometric_release": false,
  "speaker_labels": "pseudonymous_rotating",
  "pii_redaction": {"transcript": "entity_replace", "audio": "bleep_spans"},
  "sample_rate_hz": 8000,
  "codec": "G.711 mu-law",
  "revocation_status": "active"
}

Pair the record with a data statement describing speaker demographics, dialects and recording conditions [14]. If the supplier's consent came from its own customers, remember that FTC staff have warned AI companies about training on customer data contrary to their privacy commitments [12], so check the supplier's privacy notice, not only its IVR script. Sample rate also matters to fitness: 8 kHz telephony is covered in telephony vs wideband ASR training.

Clause checklist for a speech data license

Use this as a markup checklist against the supplier's draft.

1. Permitted uses enumerated per consent tier (ASR, diarization, voice agent, multi-speaker TTS, single-voice synthesis).
2. Voice Clone defined and prohibited, including zero-shot prompting and exported embeddings.
3. Speaker identification, verification and voiceprint extraction prohibited absent written biometric releases.
4. Clip-level consent manifest delivered with the data and warranted accurate.
5. Revocation mechanics: notice, purge of audio and features, and treatment of affected checkpoints.
6. Separate retention periods for raw audio, transcripts and derived features.
7. No joining of audio to identity sources; pseudonymous rotating speaker labels.
8. Output restrictions: no synthetic speech attributable to a licensed speaker in production.
9. Jurisdiction disclosure for recordings (all-party consent states, Illinois, Texas, Washington, EU).
10. Warranties and indemnity sized to biometric exposure; see data warranties.

The broader negotiation sequence sits in the AI data licensing guide, and the buyer-side view of audio sourcing is at license voice and audio data and voice agent training data. If your source is a contact center, read can I license call recordings? and contact center call recordings before scoping. To describe a specific speech dataset and its intended uses, start a request through SourceX for buyers.

Licensing voice and call recordings through SourceX

SourceX sources operational datasets, including support and sales histories, from US companies and manages the licensing process; every dataset is rights-reviewed for ownership and consents and delivered under a license defining records, uses, term and delivery. Names, phone numbers and account numbers are removed or replaced before delivery with the method recorded, though no method is perfect, and nothing is contracted until the supplying company agrees. Describe the recordings and uses you need at https://sourcex.si/buyers.