Experiment · voice cloning · 2026-10-22
Me and My Voice Clone
I trained a text-to-speech model on my own voice, then sat down and had a conversation with it. Every line on my side is a real phone recording. Every line on its side is generated. Can you tell which is which?
- 103lines: 51 real, 52 cloned
- 88 minof my voice to train it
- 1.7Bparameters (Qwen3-TTS base)
- 50 / 52clone lines right first time
Why
This is my third attempt at cloning my voice. An Orpheus fine-tune and a zero-shot commercial clone both came out American. Meorav, the text-to-speech line of my Hebrish speech project, is the first one that keeps the Irish accent. It's a full fine-tune of Qwen3-TTS-12Hz-1.7B on 1,103 sentences I read aloud, 88 minutes in all. The obvious test was to put it next to me.
On My Weird Prompts, two AI hosts normally discuss a prompt I send in. This episode drops them: it's just me and the clone, talking about how much it sounds like me.
How it was made
- Script. Claude wrote the dialogue: 103 lines, alternating between me and the clone.
- Clone lines first. Meorav v2.0 rendered all 52 of its lines on one NVIDIA L4 on Modal.
- My lines, one take. I read my 51 lines into my phone in one take, from a read sheet. I never heard the clone while recording.
- Speech-to-text. Whisper large-v3 transcribed my take with word timestamps. Aligning those words to the script found each of my lines. The same model transcribed every clone render back and checked it against the script. Two renders were wrong and were redone.
- Splice. Each clip was cut, level-matched to −20 LUFS and interleaved with 0.4 s gaps. One shared EQ and compressor went over both voices, so neither gets its own sound.
The full method, every clip, the alignment data and the scripts are in the Space.
Has anyone done this before?
Partly. Reid Hoffman interviewed a deepfake of himself, and Peter Russell interviewed his own clone. Every example I found used a commercial cloning service. I found none built on a voice model the person had fine-tuned on their own recordings, with real takes cut in by speech recognition. That probably makes this one of the first of its kind, not the first. The prior-work tab lists the examples.
Used to make it
- Hebrish/Meorav: the voice clone (v2.0, revision 600a2820; access by request, non-commercial use free)
- Hebrish-TTS-Voice-Recordings: the 1,103 training takes
- Read-Aloud-Voice-Sentences: the corpus every recording comes from
- Hebrish-ASR/Meorav-TTS: training code
- Qwen3-TTS-12Hz-1.7B-Base: the base model