AI Voice & TTS
Text-to-speech prosody, voice cloning requirements, consent practices, recording techniques, and mixing voiceover with music.
Synthetic and recorded voice have converged enough that the remaining differences are mostly decisions rather than limitations — how the script is written, what the reference recording contained, and whether anyone asked permission. These six articles cover that territory. Two are about production technique: how punctuation, sentence length and structure control the pace and emphasis of generated speech, and what a clean reference recording actually requires, which turns out to be more about the room than the microphone. Two are about cloning: what a clone inherits from its source and the four places it reliably gives itself away, and the consent and disclosure practices that belong before a model exists rather than after. The remaining two are decisions people make quickly and regret slowly — matching voice character to what the content is doing, since a delivery that carries a tutorial will undermine an advertisement, and mixing voiceover against music, where turning the music down is the fix that eventually stops working. The consent article describes practices rather than giving legal advice, and it is written for the situation most producers are actually in: holding a recording made for one purpose and considering another. Across all six, the recurring theme is that quality is decided upstream. A cloned voice cannot exceed its reference, generated speech cannot exceed its script, and a mix cannot rescue a recording that captured the room instead of the speaker. That is why the recording article is worth reading before the cloning ones, even though it looks like the least technical of the six, and why the scripting techniques matter more than the choice of engine.
All 6 articles in AI Voice & TTS
Matching Voice Character to Content Type
Voice choice is treated as taste and behaves like casting. A delivery that carries a tutorial will actively work against an advertisement, and the mismatch reads to the audience as something being wrong with the content.
Consent and Disclosure When Using a Cloned Voice
The technical question is whether a voice can be reproduced. The one that matters in practice is whether the person agreed to it, for what, and for how long — and most recording agreements never contemplated the question.
Mixing Voiceover With Music: Levels That Actually Work
Voice and music compete for the same frequencies, which is why level alone eventually fails. The working answer is a level relationship, automatic ducking, and space carved out of the music where the voice lives.
Voice Cloning: What It Needs and What It Can't Do
Cloning reproduces how a voice sounds. It does not reproduce the decisions a speaker makes about emphasis, timing and meaning, which is why a clone can be accurate and still sound wrong.
Getting Text-to-Speech That Doesn't Sound Robotic
Most complaints about synthetic speech are complaints about the script. The engine has no idea what the sentence is for, so every cue about pace and emphasis has to be written into the text itself.
Recording a Clean Reference Clip for Voice Work
A reference recording is the ceiling on everything built from it. Room reflections, distance and background noise are all captured permanently, and no processing afterwards removes them cleanly.