The audio tower · the atrium

Text to speech, voice cloning and transcription

This is the entrance to the audio tower. Text to speech that does not sound like a railway announcement, voices cloned from a few minutes of recording, music generated from a sentence, and the reverse of all of it — turning speech back into text.

Audio is where the change of the last few years is most audible, and where the ethical questions are the least abstract. A synthetic voice reading an article is useful and harms nobody. The same technology applied to somebody else’s voice without their agreement is a different act entirely, and it is the reason one wing of this tower has consent in its title rather than in a footnote.

The text-to-speech wing itself is largely good news, and unusually so. The built-in voices on every phone and computer have improved enormously, they cost nothing, they work offline, and for the most common uses — listening to an article, proofreading your own writing by ear, accessibility — they are now genuinely sufficient. The floors say that before naming a single paid service.

The transcription wing has changed just as much and is less noticed. Turning a recording into accurate text used to be a paid service with a turnaround time; it is now something a laptop does by itself, free, offline, in most languages. Those floors are mostly about accuracy, speaker separation and what to do with the result.

The music wing is the newest and the most unsettled. What may be generated, what may be published, and what the terms permit are all genuinely open questions, and those floors say so rather than guessing.

Three levels, and you are on the first. This atrium lists the floors; a floor covers one subject completely and holds the shorter guides underneath it. Nothing here is more than three clicks from the front door.

One practical warning belongs at the door, because it applies across three of the four wings and almost nobody checks it. When you upload audio to a service — a recording of a meeting, an interview with a source, a message from a relative — you are handing somebody else a recording of people who did not agree to that. Several of these services state plainly that uploaded audio may be retained or used to improve their systems, and a few say the opposite just as plainly. The difference is written down in each case, every floor here quotes it with the date, and where a free offline tool does the same job the floor names that first. For anything sensitive, offline is not a preference on these floors, it is the recommendation.

What this tower will not do

It will not help anybody clone a voice that is not theirs. Every floor in that wing leads with consent, and the floor about voice scams exists because this technology is already being used against families.

It will not sell a subscription for something your laptop does free. Both the reading-aloud and the transcription wings begin with the built-in and offline options, because for most readers that is where the honest answer ends.

It will not pretend the music rights question is settled. It is being argued in courts in several countries, the terms of the services differ from one another, and floors that touch it say what is unresolved.

And it will not carry affiliate links, including for the voice platforms, which pay recurring commissions and account for most of what currently ranks in this field.

A last note on disclosure, which runs through the whole tower. Where synthetic audio is used in something an audience will hear, these floors take the position that saying so costs nothing and settles the question before anybody asks it. What holds instead is simple: the free and offline options come first in every wing, and no floor carries an affiliate link.