Audio

Audio: how to choose

Audio is the category where machine output is closest to being indistinguishable from human production, and it is also the one where the ethical questions are most immediate. Synthetic voices are convincing enough that consent and disclosure are practical concerns rather than theoretical ones.

What actually differentiates them

For voice synthesis, prosody matters far more than timbre. Nearly every tool produces a clean voice. Fewer produce one that places emphasis correctly, pauses where meaning requires it, and does not flatten a question into a statement. Test on a paragraph with varied punctuation rather than on a single neutral sentence.

Control over delivery is the second differentiator. Being able to mark emphasis, adjust pace, insert a breath or specify an emotion is what separates a usable narration tool from a robot reading. Support for these controls varies widely.

For transcription, measure accuracy on your own audio. Published accuracy figures are produced on clean recordings with neutral accents. Real conditions involve background noise, accents, several speakers and vocabulary specific to your field, and that is where tools diverge. Speaker separation and timestamp precision matter as much as raw word accuracy.

For music, the questions are commercial rather than technical. Length, stem separation and above all what licence attaches to the output determine whether you can actually use what you generate.

What to check before committing

Voice cloning carries obligations the vendor does not carry for you. Reproducing a real person voice requires their consent, and in several jurisdictions the law is explicit about it. A vendor that lets you clone a voice from a short sample has made the act easy, not lawful.

Check whether disclosure is expected in your context. Several platforms now require synthetic media to be labelled, and some vendors embed a watermark in generated audio, which you should know about before publishing.

For music, read the licence carefully. Rights vary from full commercial use to personal use only, and some vendors reserve the right to license the same generated track to others. If the music accompanies a commercial video, this is the first thing to settle.

How pricing usually works

Voice synthesis is billed by characters or by minutes of generated audio, with voice cloning usually reserved for paid tiers. Transcription is billed by minutes of input.

The cost that surprises people is regeneration. Correcting a mispronounced name means regenerating the passage, and long form narration involves many such passes. Estimate from your expected number of regenerations rather than from the finished duration.