Text to Speech vs Voice Cloning vs Dubbing vs Transcription

Transcription takes audio in and gives text out, with timings. Text to speech takes text in and gives audio out in a preset voice. Voice cloning takes a sample in and gives audio out in a voice modelled on that sample. Dubbing takes audio in and gives audio out in another language.

The direction of travel tells you which tool you need. Audio to text: transcribe. Text to audio: text to speech or clone. Audio to audio across languages: dub.

If you need captions, subtitles or a searchable transcript, transcribe. If you need a neutral narrator for a product video or a faceless channel, use a preset voice. If the audience expects a specific person, clone that voice with consent. If the audience speaks another language, dub.

Two combinations are common in practice. Transcribe first, then edit the text, then narrate it: that is how a faceless video gets made. Or dub a finished clip, then transcribe the dub to get subtitles in the new language.

Preset voices are the most predictable, because they were trained to read clean text. Cloned voices inherit the quality of the sample: a quiet, close recording clones well, a noisy one carries the noise into every sentence. Transcription is the most sensitive to input quality, because music and crosstalk blur the words before the model sees them.

Dubbing is the least predictable, because it chains three models and any weak link shows. Clean single-speaker audio is the single biggest improvement you can make.

Transcription accepts files up to 15 MB and allows 30 per day. Text to speech allows 3,000 characters per standard request and 8,000 in long-form, with 40 standard and 10 long-form runs per day. Voice cloning works from a short sample. Dubbing accepts clips up to 3 MB and allows 15 per day.

Audio tools are much lighter than video generation, so the limits exist to keep shared capacity usable rather than to gate access. If a limit blocks real work, contact support instead of opening new accounts.

Cloning a voice requires permission from the person it belongs to, and dubbing requires rights to the recording. Preset voices avoid the issue entirely, which is one reason they are the safe default for commercial work.

Disclose synthetic speech when a listener could reasonably think a real person said it in that language. A dubbed tutorial for a new market is expected. A cloned endorsement is not.

Record or source a clean clip. Transcribe it and fix names and numbers. Translate or rewrite the text for the target audience. Generate the voice, dub it, or clone your own, then check the first and last seconds of the output.

Keep the transcript and the source text. They are the cheapest insurance against a regenerate that sounds nothing like the version you liked.