AI Voice Cloning on Mac — Useful, Creepy, and How to Do It Right

My Mac can read this article aloud in my voice. Not a voice — my voice, with my slightly Czech-flavored English vowels, cloned from fifteen minutes of recordings, generated entirely on-device. The first time I heard it, I felt two things at once: “this is incredibly useful” and “this should worry everyone.” Both reactions are correct, and any article that gives you only one of them is lying to you. So this one gives you both: the legitimate uses, the actual local tools with a full walkthrough, the honest quality verdict, and an ethics section with teeth.

The uses that justify the technology existing

Narrating your own writing in your own voice. I publish text; some readers want audio. Recording myself reading every post costs 4–5x the article’s length in retakes and editing. A clone narrates a 1,400-word post in ninety seconds of compute. The byline and the voice belong to the same person — that coherence is the legitimate core of this whole technology.

Audiobook drafts of your manuscript. Hearing your own book read back exposes clunky sentences your eyes skip over. A cloned-voice draft is a self-editing tool long before any production decision.

Accessibility — the use case that ends every cynical argument. People diagnosed with ALS, throat cancer, or facing laryngectomy can bank their voice ahead of medical voice loss and keep speaking as themselves through assistive devices afterwards. This is exactly why Apple built voice cloning into macOS itself. If you or a family member faces such a diagnosis, set this up now, while the voice is strong.

Localizing your own content. As a Czech writing in English, I can draft Czech audio versions of my English posts in my own voice rather than a generic TTS Czech. Same speaker, both languages — with open-source multilingual models this actually works, accent quirks and all.

Notice the pattern: every legitimate use clones your own voice or one explicitly entrusted to you. Hold that thought.

Walkthrough 1: Personal Voice — Apple’s consent-gated clone

macOS ships voice cloning in System Settings, and almost nobody knows. Personal Voice (macOS Sonoma and later, Apple Silicon required) builds a synthetic copy of your voice entirely on-device:

1. System Settings → Accessibility → Personal Voice → Create a Personal Voice.
2. Authenticate. Apple gates creation behind your device password — a deliberate consent checkpoint: you can’t casually build one from someone else’s recordings, because the process requires the speaker to read prompts live into the mic.
3. Read 150 short phrases aloud. Took me about 20 minutes in a quiet room; you can pause and resume.
4. The Mac trains overnight (on M-series it’s typically done in a few hours, on-device, no cloud).
5. Use it via Live Speech (Accessibility → Live Speech): type, and your Mac speaks in your voice — in calls, FaceTime, or in person. It’s also available to apps through the speech APIs, with per-app permission.

The output is clearly optimized for assistive speech: intelligible, recognizably you, somewhat flat. For its purpose — preserving a voice — it’s quietly one of the most humane features Apple has ever shipped. It is not, however, a content-production tool; for narration you want more control.

Walkthrough 2: the open-source path for narration

The open model scene moves fast; the current practical choices on Apple Silicon are XTTS-v2 (Coqui’s multilingual model — clones from a short reference clip, supports Czech among 16+ languages) and F5-TTS (newer, excellent English quality, runs well via PyTorch MPS on M-series). The XTTS path, condensed:

python3 -m venv ~/tts && source ~/tts/bin/activate
pip install coqui-tts
tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 \
    --speaker_wav my_reference.wav --language_idx en \
    --text "This paragraph is spoken in my cloned voice." \
    --out_path output.wav

The craft is in my_reference.wav: give it 30–60 seconds of clean, single-speaker audio — no music, no room echo, recorded on a decent mic at your natural reading pace. Garbage reference, garbage clone; this one file matters more than any parameter. On my M2 Ultra, generation runs faster than real time; on an M3 MacBook Pro, roughly real time — a 10-minute narration takes about 10 minutes. For long articles I script it: split text at paragraph breaks, generate each chunk, concatenate with ffmpeg, then normalize to -16 LUFS.

For Czech output I feed it a Czech reference clip and --language_idx cs. Quality is a notch below the English output but absolutely usable for drafts.

Honest quality expectations

Here’s the verdict marketing pages won’t give you. For neutral narration — articles, documentation, news-style reading — current local clones are good. Friends recognized my cloned voice in a blind test; two didn’t realize it was synthetic until I told them, though both said something felt “too smooth” in retrospect.

For emotional range, it’s uncanny valley, firmly. Ask the clone for excitement, sarcasm, grief, or a joke’s timing and you get a flat impression of those things — pitch moves, the soul doesn’t. Laughter is outright broken. Long-form fiction with dialogue exposes this within a page. Other reliable tells: occasional mangled numbers and abbreviations (write “twenty-six,” not “26,” in your scripts), breathing that’s too regular, and drift on words it hasn’t heard in the reference.

Practical conclusion: cloned voice for the 80% of content that is calm narration, real voice for anything where performance is the point. And always proof-listen — one hallucinated word in your voice is worse than a typo.

The ethics section, with teeth

This technology’s failure modes aren’t hypothetical: voice-clone scam calls impersonating family members are documented and growing, and a cloned voice has already been used in election robocalls. So here are the bright lines as I practice them, stated as rules, not vibes:

Consent is the bright line — full stop. You clone your own voice, or a voice whose owner gave you written, informed permission for a specified use. Not your late grandfather from family videos without the family’s agreement. Not your favorite streamer “as a joke.” Not your colleague for the office party. The technical ability to clone from 60 seconds of audio means everyone you’ve ever recorded is cloneable; the only thing standing between that fact and abuse is the norm this paragraph states. Apple understood this — Personal Voice’s design makes the speaker’s live participation mandatory. Treat that as the moral spec, even when open tools don’t enforce it.

Disclose AI narration. Every audio version of my posts ends with one sentence: “This narration was AI-generated from my own voice.” It costs five seconds and keeps trust intact. Listeners forgive synthesis; they don’t forgive discovering it.

Know the legal weather. Deepfake and voice-replica laws are emerging fast — the EU AI Act imposes transparency obligations for synthetic media (deepfakes must be labeled), several US states now have voice-likeness statutes (Tennessee’s ELVIS Act explicitly covers voice), and impersonation for fraud is criminal everywhere. “It’s just AI” is not a defense anywhere, and the direction of travel is stricter.

Protect your own voice too. Agree on a family verification question for urgent “emergency” calls asking for money. That advice used to sound paranoid; it now appears in police-prevention leaflets, including Czech ones.

Where this leaves us

Set up Personal Voice this weekend — genuinely, even just as insurance; it’s twenty minutes of reading and you’ll possess something irreplaceable if you ever need it. If you publish, try the XTTS path on one article and see whether your readers want audio. And hold the line: your voice, or written permission, disclosed when used. The tools are now easy enough that the ethics are the only hard part — which means the ethics are the skill.

Useful and creepy were never opposites here. The same fifteen minutes of training data serves a patient banking their voice before surgery and a scammer rehearsing a fake emergency call. The technology doesn’t choose between them. The person at the keyboard does — do it right.