Me and My Voice Clone

Also on YouTube. The same video is in this Space as an MP4 download.

The video shows who is speaking: orange is me, blue is the clone. Close your eyes for the first minute and see how long you can tell them apart.

The episode as published

This is the full episode as it ran on My Weird Prompts, with the show's jingles and disclaimer. The dialogue alone, without the jingles, is dialogue-only.m4a (8:12).

The idea

I've tried cloning my voice before. An Orpheus fine-tune and a zero-shot commercial clone both came out American. Meorav is the third attempt, and it keeps the Irish accent. The obvious test was to sit it down next to me and see whether anyone can tell which of us is which.

Normally on My Weird Prompts I send in a prompt and two AI hosts, Herman and Corn, discuss it. This episode drops them. It's just me and the clone, talking about how much it sounds like me, what it gets wrong, and what it's like to hear yourself say things you never said.

Every line of the conversation as a separate clip, exactly as it went into the episode. Play them in any order to compare.

How it was made

The clone can't hear me. The whole conversation was written in advance, and the two sides were recorded separately and spliced together.

1. Script. Claude wrote the dialogue: 1,710 words, 103 lines, alternating between me and the clone. (script)
2. Clone lines first. All 52 clone lines were rendered by Meorav v2.0 (Hebrish/Meorav at revision 600a2820) on one NVIDIA L4 on Modal: 280 seconds of speech in 728 seconds of compute.
3. My lines, one take. I read my 51 lines from a read sheet on my phone, in one 5:41 take, pausing about three seconds between lines. The clone's lines were on the sheet in italics, for timing only. I didn't hear them while recording. (read sheet, raw take)
4. Speech-to-text to find my lines. My take went through Whisper large-v3 (faster-whisper, int8, CPU) with word timestamps. Its words were aligned to the script with a sequence match, which gave a start and end time for each of my 51 lines. (words, spans)
5. Speech-to-text to check the clone. Each clone render was transcribed back with the same Whisper model and compared with its script line. Two had gone wrong. One dropped its last sentence. The other said "I never" instead of "I'd never". Both were re-rendered. The other 50 were correct first time. (check)
6. Splice. Each of my lines was cut from the take with 0.15 s of lead-in and 0.25 s of tail and short fades. Silence was trimmed from the clone clips. Every clip was level-matched to −20 LUFS, and the 103 clips were interleaved with a 0.4 s gap between turns.
7. One shared chain. Both voices then went through the same EQ and light compression together, so neither sits in its own acoustic space: an 80 Hz high-pass, an 11 kHz low-pass, −2 dB at 250 Hz, +2 dB at 3 kHz, and a 2.5:1 compressor. The 8:12 dialogue was then published through the normal My Weird Prompts pipeline, which adds the jingles, a disclaimer and mastering.

The scripts are in pipeline/: asr.py (Whisper word timestamps), align.py (script alignment), check_clone.py (round-trip check), build.py (cut, level-match, stitch, EQ) and render_video.py (this video). Each was run by hand in this order, not as one script.

What gives it away

• The microphone. The clone learned from quiet-room takes on a good microphone. My side is a phone, so the clone is often the better-recorded of the two.
• The delivery. The training data is read sentences, so the clone sounds like me on a good day: never tired, never looking for my keys. The episode makes this point itself.

There was no formal listening test. That's the next step.

Used to make this

Has this been done before?

Partly. People have talked to clones of themselves before. Reid Hoffman interviewed "Reid AI", and Peter Russell interviewed his own clone. Every example we found used a commercial cloning service. We found none that used a voice model the person fine-tuned themselves on their own recordings, with a scripted dialogue and real phone takes cut in by ASR alignment. That doesn't make this the first. It means we didn't find an earlier one.