September 11, 2026. On 9 September Twilio moved ElevenLabs voices for the <Say> TwiML verb into public beta. You select one with voice="ElevenLabs.<Voice ID>", and that is the whole integration. What you do not select is the model. Twilio picks it for you from the language attribute, and the rule has a consequence for every deployment that is not US English.
The language attribute picks the model
English locales are routed to ElevenLabs Flash 2 and every non-English locale is routed to Flash 2.5. Twilio's own wording is that custom model selection is not currently supported, so a call in Spanish and a call in English from the same application run on different speech models whether or not you intended that. Worth knowing if you are matching a voice a client signed off on: Twilio spells the model three different ways across its own surfaces, as "Flash v2 and Flash v2.5" in the changelog, "Flash 2.0" in the auto-selection description, and "Flash 2 and Flash 2.5" in its AI Nutrition Facts label.
Pronunciation control stops at en-US
This is the part that will bite an agency first. Twilio's Text-to-Speech documentation restricts SSML support for ElevenLabs voices to the en-US language. Twilio's ConversationRelay voice page states the same restriction while listing a shorter set of supported tags, so the two pages differ on exactly which tags survive but agree on the boundary.
Read that as the operational rule rather than a documentation quirk. SSML is how you force a pause, fix a mispronounced surname, or make a policy number read as digits instead of a quantity. Outside US English, on ElevenLabs voices, you do not have that lever. Note the boundary is en-US, not English, so a British or Australian deployment loses it too. If you are running AI calling for insurance, where a bot reads back names, policy numbers and dates all day, that is the difference between a call that sounds competent and one a customer hangs up on.
The voice is not contractually stable
Twilio is unusually candid here, and it is in the legal notice rather than the changelog. Third-party voices "might change without prior notice", and Twilio spells out that this includes "new models that affect how voices sound or the removal of voices from their offering with or without alternative or automatic redirections". Separately, Twilio states that its Service Level Agreement does not cover Beta products, and this is a public beta.
Billing is straightforward by comparison. Twilio rounds text-to-speech usage towards the end of the call and prices it in blocks of 100 characters, with the minimum charge covering 100 characters.
What it means for operators
Three things to do before you put an ElevenLabs voice in front of a client's customers. First, test in the client's actual locale, not in en-US, because the model and the SSML support both change underneath you. Second, if pronunciation accuracy matters and the deployment is not US English, evaluate the Amazon Polly and Google generative voices on the same script before committing, since they carry different SSML support. Third, record which Voice ID you shipped and re-test it on a schedule, because Twilio has told you in writing that the voice can be changed or withdrawn without notice and that beta sits outside the SLA.
One piece of useful context on maturity: ElevenLabs voices reached Twilio's ConversationRelay back in March 2025, and only now reach <Say>. The two surfaces even take different voice values, {Provider}.{Voice} for <Say> against a bare {Voice} for ConversationRelay, so configuration does not port between them cleanly. If you are wiring voice into a wider client stack, that plumbing is the part worth planning as AI automation rather than treating as a one-line attribute change.
Frequently Asked Questions
No. Twilio selects the model from the language attribute on your Say verb: English locales are routed to Flash 2 and non-English locales to Flash 2.5. Twilio states that custom model selection is not currently supported.
Only for en-US. Twilio's Text-to-Speech documentation restricts SSML support on ElevenLabs voices to the en-US language, and its ConversationRelay voice page states the same restriction with a shorter tag list. The limit is the en-US locale specifically, so en-GB and en-AU deployments lose it too.
Twilio rounds usage towards the end of the call and prices it in blocks of 100 characters, with the minimum charge covering 100 characters. Basic voices are free; standard, neural and generative voice rates are on Twilio's voice pricing page.
No. It is a public beta, and Twilio states that its Service Level Agreement does not cover beta products. Twilio also warns that third-party voices may change without prior notice, including new models that alter how a voice sounds and the removal of voices altogether.