The idea
Every episode of My Weird Prompts starts with Daniel sending in a prompt, and the AI hosts take it from there. This one has no hosts. It is Daniel talking to his own voice clone about how much it sounds like him, and the clone giving as good as it gets: it boasts about its 1,103 training takes, delivers an Irish tongue twister without trouble (it has no tongue), and points out that it can't laugh, whisper or shout, because he never did any of those in the recordings it learned from.
The twist is in the timing. The clone's lines were written and rendered before Daniel recorded his, so it can't hear him. It just knows what he's going to say. Listen for the game near the end, where both say the same sentence and the listener is asked which one was real.
How the voice clone is made
Meorav (Hebrew for “mixed”, as in the Jerusalem mixed grill) is a full fine-tune of Qwen3-TTS 1.7B on Daniel's own recordings. It is part of Hebrish, his project on English speech with Hebrew words mixed in.
- The data. 1,103 read takes, about 88 minutes after trimming: phonetic sentences that cover every English sound, conversational lines in his own register (with some Hiberno-English, because reading aloud pulls towards a neutral voice), a large set of English sentences with Hebrew words in them, names of places, and some personal lines.
- The training. Every weight of the base model is trained, not a small adapter: learning rate 7×10-6, three epochs, on one rented GPU.
- Why not a quick clone? Clones made from a few seconds of audio, which is how the show voices everyone else, keep the timbre and lose the accent. Daniel's came out American. A full fine-tune on an hour and a half of his speech is what moves the accent.
- Hebrew in English sentences. Because the training sentences mix the two, Meorav says words like yalla, tachles and Bituach Leumi the way he does, in the middle of an English sentence, with no separate Hebrew engine.
The episode used v2.0. The same evening v2.1 replaced it: the same recipe on the same takes, cleaned first with DeepFilterNet 3 to remove room noise. Both are published as Hebrish/Meorav (access on request, licensed CC BY-NC 4.0).
How the episode was put together
- The script. 51 lines for Daniel, 52 for the clone, about 1,600 words, written with Claude.
- The human half. Daniel read all 51 lines in one take on his phone, leaving a pause after each. Silence detection split the 5 minutes 41 seconds into exactly 51 pieces, and a speech-recognition pass confirmed each piece matched its line.
- The clone half. Meorav rendered the 52 lines on a cloud L4 GPU in about twelve minutes, at roughly 2.6 seconds of compute per second of speech.
- Checking the clone. Every clone line was transcribed and compared with the script. Two had slipped: the longest line, 50 words, lost its last sentence (“I'd like you to tell the listeners which one of us is the voice clone”), and “I'd never” came out as “I never”. Both were re-rendered and spliced in. v2.1 renders the long line in full.
- Making two sources sound like one room. A phone microphone and a speech model don't sound alike. Each clip was set to the same loudness, then one EQ ran over the whole conversation: a low cut at 80 Hz, a high cut at 11 kHz (the model's output stops at 12 kHz, so the phone's extra top end was the giveaway), a little less mud at 250 Hz, a little more presence at 3 kHz, and gentle compression. The show's usual room tone and mastering went on top.
- Publishing. The pipeline normally writes and voices every episode itself. This one arrived finished, so it went out through a new route that takes a ready-made dialogue track and skips speech generation, while the jingles, cover art and publishing run as usual.
Where the clone lives now
Meorav isn't a one-off. Since 22 October 2026 it reads the AI disclaimer at the top of every episode and the introduction to each listener's prompt, and the pipeline picks up each new version of the model as soon as it is published. The hosts still run on Chatterbox, a fast general-purpose engine. The technical white paper has the details.