XTTS-v2 is regularly described simply as open source. That phrase hides an important distinction: its model checkpoint is governed by a specific public model license rather than a blanket permission to monetize outputs.
Version and model identity
Coqui’s model catalog identifies XTTS-v2 under a multilingual model name and labels its checkpoint license CPML. The project’s XTTS documentation describes cross-language speaker conditioning and lists supported languages. Model names alone are not immutable endpoints; record the model config, checkpoint checksum and exact release if building a repeatable voice project. A README can describe capabilities without establishing rights to an individual’s voice.
Source license versus weight license
Software code and model weights can have different rules. A code package with a permissive license does not supersede the model’s Coqui Public Model License. Teams need the actual license text controlling the checkpoint, not a blog summary or third-party inference-service terms. If the proposed use is monetized content, client voice cloning or sublicensed distribution, set a legal review gate and keep commercial authorization unknown until the license has been evaluated.
What cross-language cloning gets wrong
A short sample contains timbre, accent, background noise and microphone signature. A target-language synthesis can resemble the speaker but mispronounce proper names or carry source-language phonetics into another language. Make native-language review mandatory for prominent brands and sensitive terminology. Text cleanup should distinguish spelling out an abbreviation from changing its meaning, and a transcription of a source sample should not be silently assumed correct.
An auditable local production run
Keep the original consent document separate from the text-to-speech job. Store checkpoint hash, locale tag, speaker reference ID, decoded text, generation settings and version. Listen for clipping, unintended utterances and inconsistent names across long chunks. Add an expiration or deletion process when consent ends. The model running offline does not erase a performer’s likeness rights or the obligation to protect their recording.
Decision alternatives
Use Piper for non-cloned offline announcements that can tolerate less expressive narration, Chatterbox for a separately validated open-weight generative workflow or a licensed studio service when actor agreements are central. XTTS-v2 remains a legitimate technical candidate but should not be described as guaranteed freely commercial without evidence from the governing model license.
Publication scope, rights and cost caveats
Research cutoff: 25 October 2026. The material above distinguishes provider documentation from independently measured behavior; no hands-on evaluation, legal clearance, voice-owner consent, commercial license purchase or model benchmark is claimed. Production teams must verify the specific API or installed version, account entitlement, input/output rights and destination market before release. Monetary plan amounts and metering tiers are not asserted here; the canonical pricing ledger belongs to AI 10. Missing release dates, exact rates and sample-rate capabilities are intentionally not inferred.