When a synthetic voice is retired: what happens to your audio

Most publishers treat a synthetic voice as a permanent asset. You audition a few, pick one, and it becomes the sound of your site. Listeners get used to it. Some newsrooms even give it a name. What rarely comes up during the trial is that the voice is not yours to keep: it lives inside a model run by someone else, and that someone can withdraw it.

We found this out the practical way in early November 2026, when one of the voices used by our customers stopped existing on the provider's side. This piece is what we learned, written as the three questions worth asking before it happens to you.

Question one: does your back catalogue survive?

It depends on a single design decision, and it is worth checking with any vendor. Some tools generate audio on the fly, at listening time, or regenerate it on a schedule. In that model, a voice that disappears takes your archive down with it: every play becomes a failed request.

WeDispatch does the opposite. An audio is generated once, stored as an MP3 file, and kept version by version. Playing it never calls the voice again. When a voice is withdrawn, nothing already published changes: the player, your podcast feed and any listening links you have sent keep working exactly as before. This is also why we frame portability in terms of files, as explained in switching audio provider without losing your audio.

One caveat. If you regenerate an old article after the voice is gone (to fix a typo, say), the new version is read by a different voice. On a standalone story, few listeners will notice. On a multi-part series meant to be heard in one sitting, they will.

Question two: what happens to the next article?

This is where tools differ, and where we got it wrong the first time.

Our pipeline had a sensible rule: voice-related errors from the speech engine are almost always transient (timeouts, rate limits, a busy server), so the job is retried. A "voice not found" response fell into that bucket. The result was a job that retried once a day, for three days, against a voice that would never answer, while the customer's dashboard advised patience. It was reported to us on 5 November.

The fix shipped the same day, in version 1.166.1. A "voice not found" response is now treated as final. The job fails at once instead of retrying, because retrying with the same voice will produce the same answer. The reason shown in the customer's audio list says what is actually true: the chosen voice is no longer available, pick another one on the Voice screen, then regenerate. And the failed attempt is not charged: nothing is taken from your credits.

Question three: what does the catalogue do?

Our voices are not a hand-maintained list. They live in a registry where each voice carries three dates: when an automated survey first saw it at the provider, when a person on our team listened to it and approved it, and when it disappeared.

When the survey finds that a voice is gone, it does not delete the entry. It dates it. Keeping the line means an old voice identifier still stored in a customer account can be explained months later. But a dated voice leaves the list customers choose from, and, since 5 November, it is also rejected everywhere a voice is validated before reading. That second part was the missing piece: the retired voice had left the menu, but its identifier could still slip through the path used for cloned voices, so every job using it kept failing.

For an account whose default voice was retired, the next articles are simply read by the default voice for the article's language. No action is needed for publishing to continue. That fallback is a safety net, not an editorial choice, so if a specific sound matters to you, go to the Voice screen and pick a replacement deliberately, after listening, as suggested in our French text-to-speech guide.

Cloned voices are a special case

A signature voice cloned from your own presenter belongs to your account, not to the shared registry, so we cannot know in advance that it has disappeared. The immediate-failure rule still applies: the first job that asks for it fails straight away, with no retries, no charge, and a plain reason. You then either recreate the voice or choose another. What a signature voice is, and what it commits you to, is covered in signature voice: what the term means for publishers.

A short checklist for any vendor

• Are published audios standalone files? If playback depends on the voice still existing, your archive is only as durable as the model behind it.
• What exactly happens when a voice vanishes? Look for described behaviour (immediate failure or retries, the message shown, automatic fallback or not), not a promise to "let you know".
• Who pays for failed jobs? An attempt against a voice that no longer exists should cost you nothing.

Our own answers are the ones above, including the unflattering part: the first time this happened, it took us three days to handle it properly. Both safeguards, rejecting retired voices and failing without retries, are now covered by an automated check that turns red if either one is removed from the code.

Read next

• Quote, unquote: should a text-to-speech article announce its quotations?
• Changing your audio volume mid-subscription: what happens to your credits and your invoice
• French text-to-speech for children: choosing a voice for young listeners