What does it take to teach a machine to say Bituach Leumi the way somebody who actually lives in Jerusalem says it? That's the heart of Daniel's prompt. He trained a text-to-speech model on his own voice, and the voice you hear every time he speaks in this episode is that model. Daniel, you're in the room.
I am, but I'm technically not speaking at all. I typed every word of this, and a model called Meorav is saying them. I'm the only person in the studio who isn't talking.
Which makes this a side-by-side. Corn and I run on Chatterbox, the engine the show uses for all of us. Daniel runs on Meorav. And one thing worth saying early, because listeners will wonder. The Daniel voice you're hearing is version two of Daniel's voice model. The second Meorav. The first was an experimental proof of concept, and it's already published on Hugging Face under the Hebrish organisation as TTS V one.
Version two. So version one is the one we're not meant to judge too harshly.
It ran end to end, and that's what a proof of concept is for. Version two is the one Daniel wanted to put next to the two of us.
Fine, but go back a step. Meorav. It sounds like something you'd order at a counter.
It is, sort of. Meorav is Hebrew for mixed, as in meorav Yerushalmi, the Jerusalem mixed grill. Different cuts, one grill, everything touching. I liked that for speech that mixes English and Hebrew in the same sentence.
And it lives inside Hebrish, which is Daniel's one-person not-for-profit on English and Hebrew code-switching. The speech recognition models there are named after Israeli food too. Sabich, Jachnun, Shakshuka.
Of course the recognition models are breakfast. What's code-switching look like in practice?
It's how English speakers talk here. You're halfway through an English sentence and a Hebrew word walks in and nobody blinks. Yalla, tachles, Bituach Leumi wants the form by Thursday.
That's the sentence the whole project is built around. Irish-accented English, because that's Daniel's accent, with the Hebrew words said the way he says them, in the middle of the sentence and not as a separate performance.
So why not just clone yourself from a clip, like everybody else? We've done cloning from a few seconds of audio on this show, and it works.
It works, and I tried it. Twice, actually. An Orpheus fine-tune and a zero-shot clone. Both came out American. It sounded like me if I'd spent a summer in Ohio.
There's a mechanism behind that, and it's worth a minute. A zero-shot clone copies timbre, the colour of the voice. But the rhythm, the stress, where a sentence rises and falls, all of that comes from the base model, and the base model learned mostly from American speech. So you get your voice reading in somebody else's cadence.
Your voice wearing a stranger's coat.
And the show has its own example. Daniel's voice already exists here as a Chatterbox clone, and it reads the AI disclaimer at the top of every episode. A single-shot clone, so the timbre is his, and the cadence is less Irish than the real thing.
I can confirm that. Hearing yourself read a legal disclaimer in a mild American accent is a particular kind of unsettling.
Worse than hearing yourself normally?
Different. Normally I hate the pitch. With that one I hate the vowels.
Corn, you've used the clone for a lot of things. Does it sound like him to you, or does it sound like a very polite American doing an impression of him?
The second one. It's missing the lilt. Daniel says the same sentence and it goes up at the end like a question, and the clone lands it flat like a receipt.
That's the prosody coming from the base model. The fix is to stop borrowing and change the model itself. A full fine-tune updates every weight, so it can move off the American prior rather than being nudged around it.
You said that like it's settled. Is it?
No. I haven't seen a published open text-to-speech fine-tune that holds a non-American English accent. Not one. The project's own notes say so. Whether it works is the thing being tested.
It might not. That's the project.
A bet that admits it's a bet. What did you build it on?
Qwen three TTS, the one point seven billion base model, from Alibaba's Qwen team. It's Apache licensed, so I can fine-tune it and publish what comes out.
Qwen three TTS, twelve hertz, one point seven billion parameters. And the detail that made it the pick is that it reads raw text. There's no phonemiser in front of it. Most speech systems turn your sentence into phonemes first, and the usual tool for that is espeak, and espeak has no Irish-English voice. So that front door was already shut.
And the tokenizer just eats the spelling.
It eats the spelling. Which means the spelling of a Hebrew word matters enormously. One spelling per word, so the model learns that this exact string means Daniel's pronunciation. He keeps a canonical list for that, shared with the recognition work.
How many ways are there to spell tachles?
More than you'd think. Takhles, tahles, tachles. People write all of them in messages. The list picks tachles, and every sentence I record uses that one.
Corn, would you have guessed that one?
I'd have guessed the h and been confidently wrong. Same as everything else.
That's also why the sentences look odd. Numbers are spelled out, the way you say them. Two thousand and twenty-six, half four, twelve euro fifty. The model maps text to sound, so the text has to say what I said.
Which is why nobody in this script has written a digit. I've been told.
Corn, can I do the training script? I've been holding it since the opening.
Go on. I can see it in your ears.
Qwen ships a fine-tuning script, and it has bugs. Daniel patched seven things. The model was learning one frame too far ahead. It was predicting codes from a hidden state that had already seen them. A projection layer was skipped in training and then used at inference. And in half precision, plain weights lose most of the small updates at these learning rates, so he trained with full-precision master weights.
The last one sounds like my week. Small effort, then it vanishes.
When Daniel last checked, none of the fixes had been merged upstream.
I'd love to say I found them cleverly. A lot of it was people on the issue tracker saying this looks wrong, and me saying yes, it does.
What does one of these runs cost?
The short ones are about a dollar. One rented GPU on Modal, a few passes over the data at a small learning rate. The first proof of concept was three passes.
Which is why starting small was sensible. You learn whether the whole pipeline works for the price of a coffee, before you give up a weekend to recording.
It depends on the coffee.
In Jerusalem it depends on the neighbourhood.
Where does the audio come from? Eight hours of the phone book?
A recording app on my phone. Raw audio, forty-eight kilohertz, twenty-four bit, browser voice processing off. And it has a voice-clone mode, which means a quiet room.
That's the clever part. The same app also collects speech recognition data, and that wants the opposite. Recognition wants an air conditioner and a crying baby. A voice model wants silence. So only takes recorded in voice-clone mode, with no noise flags, ever reach Meorav's training set.
A one-way door.
A quiet take can help the recognition side. A noisy recognition take never trains the voice.
And what do you read?
Sentences. A phonetic set so every English sound turns up. Conversational ones in my own register, with a bit of Hiberno-English, because reading aloud pulls you towards a neutral reading voice. A big set of Hebrish sentences covering the common Hebrew words. A set for names and places, Yerushalayim, Chevron, that kind of thing. And a personal set.
The personal set has the show in it. Daniel's name, Hebrish, Herman Poppleberry, Corn.
So he read my name into a microphone and nobody asked me.
You weren't available for comment.
I'm a sloth. I'm always available. I just answer slowly.
Version two is more of the same, and a lot more of it. Eleven hundred and three takes, about eighty-eight minutes after trimming, against fifty minutes for the first. The longer plan is two and a half to three hours, with a different kind of recording in it. Free speech, where he talks unscripted. That's meant to be the strongest lever on accent, about a third of the final hours.
Reading aloud makes you sound like a man reading aloud. When I just talk, the Irish comes back.
What does Irish even mean to a model? Where does the accent live?
In small things. The project audits its recording sentences for them. The th in this and thing, the r after a vowel, the diphthongs, the ing endings, and the t. Irish English as a rule keeps the r, softens the t, and makes some vowels purer than an American would.
Diphthongs. A sloth hears that word and goes to sleep.
A read sentence makes me self-conscious, so I enunciate like a newsreader. Free speech catches the sloppier, more Irish version.
Daniel, what's it for? Besides this episode.
Mostly me. News briefs I prepare for myself, read in my voice. Narration. Anything where I'd otherwise have to sit down and read a long text out loud, and I'd rather not. The disclaimer, maybe, once it's good enough.
Why your own voice and not a hired actor's?
Because I can record as much as I like, for free, and nobody else's consent is involved. I can publish the recordings because they're mine.
That sidesteps the question everyone else in this field is stuck on, whose voice it is. Here the answer is one man, and he's in the room.
Herman, a real question. We're Chatterbox and we're fine. Why would anyone want a model that needs hours of one man's voice?
Because Chatterbox is a general engine. Give it a short reference and it speaks in that timbre, fast and cheap. A whole episode runs on the order of fifteen cents of GPU. But it reads romanised Hebrew as English. It sees the letters on the page and reads them the way an English reader would. Chatterbox was actually on Daniel's fallback list, behind CosyVoice three.
Is there anything we do better than Meorav?
Plenty. Chatterbox takes a reference clip and works the same afternoon, in nearly any voice. Meorav needs recordings, a training run, and a model on a disk, and then it only does one voice. Daniel's trade is breadth for depth.
A clone is a photograph of the voice. The fine-tune is trying to learn how I move.
So I'm the plan C.
You're plan C for Daniel and plan A for everyone who just wants a donkey and a sloth.
Herman, I've got one for you. If Meorav nails the Hebrew and loses the Irish, would you call that a win?
A half win. Daniel would call it a Hebrew speaker with an American accent, and he's heard that one already from the clones. I'd want both.
Hm. So now we find out. Do we do the demo?
We should do the demo, and I'd like it clear what we're doing. Daniel says a few everyday lines, the way English speakers here really talk. Then Corn and I try the same lines, and listeners compare.
And I'll say now that it may go badly for us.
It may go very badly. I'm from Connecticut. You're from, well.
Mongolia. A very real place, and I take offence.
My guess for tachles is a tack followed by a less. That's what the letters say to an English reader.
Tackless. I've been called worse.
Right, line one is the one from the project page. Yalla, tachles, Bituach Leumi wants the form by Thursday.
Yalla is let's go. Tachles is get to the point.
Tachles is more the bottom line. And Bituach Leumi is the National Insurance Institute, the office you can't ignore.
I'll go first. Yalla, tachles, Bituach Leumi wants the form by Thursday.
Brave, Corn, or at least confident. I will leave the judging to the listeners.
Daniel, give us another before I lose my nerve.
Sababa, I'll sort the arnona on Sunday, but the office is a total balagan.
Sababa is fine, great, no problem. Arnona is the municipal property tax. Balagan is chaos.
Balagan is a mess, but it's the specific mess you get when nobody's in charge.
My turn. Sababa, I'll sort the arnona on Sunday, but the office is a total balagan.
You looked proud of yourself.
I'm a donkey. It's the only face I have.
Next one. Ma nishma, are we doing a tiyul on Friday, or is it going to be a chamsin?
Ma nishma is what's up. Tiyul is a hike. Chamsin is a heatwave, the dry wind.
That's it. Ma nishma is how you'd start with someone at the bus stop.
Ma nishma, are we doing a tiyul on Friday, or is it going to be a chamsin?
I'm going to say something kind, Corn. You committed.
I'm a sloth. Committing is all I do. It's moving I struggle with.
Two more. Take kvish echad, there's no traffic after motzei shabbat.
Kvish echad is Highway One, the road up to Jerusalem. Motzei shabbat is Saturday night, once Shabbat is out.
And this one's a city. Yerushalayim in the rain is sababa, honestly.
Yerushalayim, which is Jerusalem, the long way round. Yerushalayim in the rain is sababa, honestly. I honestly can't tell from the inside whether that was good or bad.
Neither can I. That's what the listeners are for.
What I'd listen for, if I were a listener. The ch in tachles and Bituach, which is a throat sound and not the ch in church. Whether Yerushalayim keeps all its syllables. And whether the stress lands where a Jerusalemite would put it.
That's a lot of ch for a sloth.
One more, because I've been saying nice easy ones. Slicha, I'm walking through Shuk Hacarmel to get tomatoes for the shakshuka.
Slicha, I'm walking through Shuk Hacarmel to get tomatoes for the shakshuka.
I'll say one thing. I picked those because I say them all day, which is exactly why they make a weak test. A model can learn the lines it trained on. The harder test is a Hebrew word I haven't recorded.
That's the out-of-distribution question. The project holds out sentences the model never trained on, and some lines made up to be strange, and compares the base model, the fine-tune and the real recordings.
And who judges?
My ears, and other people's. There are no metrics here. It's evaluated by ear.
Which is right. Loss going down tells you the model is learning something. It doesn't tell you it sounds like Daniel. Version one's loss came down and every sample stopped cleanly at about the length of the real takes. Nobody has said how it sounds. That verdict was still pending.
A voice that knows when to stop. Unlike this show.
Unlike your tangents. Corn, what do you think version two has to prove? Be honest.
That the Irish survives. And that the Hebrew comes out right in the middle of an English sentence. They may fight each other, Hebrew vowels being a long way from Irish ones. If it picks one, I want to know which.
I'd bet on the Hebrew. There's a lot of Hebrew in the training sentences. The accent is the harder one, and it's one speaker, about an hour and a half of read sentences this time, and version one was about an hour.
Daniel, what makes you nervous about this?
That I'll be the worst judge in the room. Nobody hears their own voice from outside. I'll need other people's ears, and I don't have a number to hand them.
Fifty minutes. Six hundred and fifty-seven takes, fifty minutes after trimming. Version two, eleven hundred and three takes, eighty-eight minutes.
Fifty. Noted.
Corn, you were right about the flat landing. The clone says the line like a receipt. A six-second clone does the job for nearly everything. An hour and a half in a quiet room is a lot of effort to get an accent right.
For most uses a clone is the right answer. I'm not saying anyone else should do this.
The licence is non-commercial. CC BY-NC. So the show can't sell it and the recordings are public. Anyone can download your voice. The people who'd misuse it don't read licences.
True, and I knew that when I published them. The alternative is me deciding who gets to read my own recordings, and I like that less.
Two licences on the one voice, too. The base model is Apache. What Daniel made from it is non-commercial. Tennant would have wanted a form for each. And once it says whatever you type, it says whatever anyone types. That one I'd take up with the council.
The council?
Nineteen seventy-three. I did the lift announcements at the Drumlish District Records Office. Third floor, births. Fourth floor, deaths. Fifth floor, the canteen, which was a cupboard. Recorded on a reel of the Corporation's tape, in the Corporation's time, so the Deputy Registrar, a Mr Aloysius Tennant, ruled the voice was Corporation property.
So you were a recording on their tape.
I needed a licence to say the word floor in public in the county. Form eleven B. Two pence a floor. Mr Tennant filled it in for me, the pen not being made for claws, and kept the carbon.
Two pence on five floors is ten pence a day. Before you've said anything else.
Twelve. The canteen was four. He said it was a different floor for tax purposes. Then he leased the reel to Ballyknock for their lift. Two floors. A goat on the payroll as second signatory, Bernadette, with authority up to forty pounds. For eleven Thursdays my voice announced a floor that wasn't there. A man went up in it looking for births and got a field.
That's what happens when you apply a model outside what it was recorded for. It had no data for a floor that doesn't exist.
It had the word floor. That was all it needed.
And the goat signed off on it.
Twice. Then I appealed to the Surveyor's Department, which ruled the voice a fixture, like a radiator, and sent a man with a spanner.
In property terms a fixture stays with the building when it's sold. So by that ruling, your voice came with the lift.
Nine pounds, with the handrail. Tennant had it on the inventory between the fire bucket and the clock. The man with the spanner was a Mr Dolan. He listened to me say floor, four times, and signed it as tightened.
A voice that needed tightening.
I did it again myself on a Saturday. One take, nine minutes, no Corporation tape. Mr Tennant played the new one in the lift for eleven years and signed for it every Monday.
Nine minutes. Version one was fifty minutes of recordings and it ran. Nine minutes and a Saturday is a start.
A voice nobody could keep, replaced in nine minutes by the man who owned it. Daniel's version two had an hour and a half of him reading. The next one is where we find out how the Irish holds up when he's talking and not reading.
Even a good version only settles one speaker. Whether it generalises is a question for version three.
One problem at a time. Thanks to Hilbert Flumingtop for producing this one.
This has been My Weird Prompts. If you want more, find us at my weird prompts dot com.
See you tomorrow.
