AssemblyAI TTS-1 separates voice choice from controls still on its roadmap

AssemblyAI’s [October 27 announcement](https://www.assemblyai.com/blog/text‑to‑speech‑api) makes AssemblyAI‑TTS‑1 generally available, with library voices and instant voice cloning. Cloning requires a recording and confirmation of the speaker’s permission. A voice choice identifies who the speech should sound like. A delivery control specifies how a particular line should be spoken. Those are different requirements for an existing voice workflow: finding a suitable voice does not establish that every control in a previous script has an equivalent. For a team considering a cloned voice, confirming permission is also a distinct step from selecting an audio sample. The announcement describes the product’s requirement; it does not establish the permissions for a particular person’s recording. OpenTools did not upload a voice, create a clone or assess a customer’s rights.

The [launch FAQ](https://www.assemblyai.com/blog/text‑to‑speech‑api) lists English, Spanish, French, German, Italian and Portuguese. Fine‑tuned Pro cloning is not supported at launch. A useful preparation record can separate the language of a request, the intended voice and the exact text that must be spoken. That keeps a language selection from becoming an assumption about how every name, abbreviation or mixed‑language passage will be rendered. For example, a team moving an appointment reminder could mark which parts of the message are fixed and which come from a customer record. It can then identify the names or reference strings that need review. That is a proposed review task, not an output we generated or a claim that the model handles every such case.

The [control limits](https://www.assemblyai.com/blog/text‑to‑speech‑api) are separate: pronunciation dictionaries and emotion tags are coming soon; SSML and a speed parameter are not accepted. The company describes punctuation and spelling as ways to shape delivery. A workflow that relies on structured pronunciation instructions or a dedicated speaking‑rate field therefore has a concrete migration question: which parts can be represented in the currently accepted input, and which depend on a future control? A roadmap item cannot answer that question for a request sent today. Changing written text to influence delivery also changes the material given to the speech model. A team can retain the original message beside any adapted version so a reviewer can see what changed before it reaches a caller. Whether that adaptation produces an acceptable result still needs to be established in the team’s own evaluation. This report distinguishes announced voice options from control availability. It does not compare model quality, latency, accuracy, call outcomes or cost, and no AssemblyAI API, account or generated audio was tested. Actor Yves Corbeil studies a script during a voice‑recording session for the animated film L’Entrevue, May 20, 2010. Archival context for preparing spoken text, not AssemblyAI use, a cloned voice, consent to cloning or an observed model test. Photo: [Kat Baulu / Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Yves_Corbeil_voice_recording_L%27Entrevue.jpg), [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/). Cropped for display.