Why I built this, 25 hours of recordings and the two models I'm keeping

This is the recording written up to be read: everything I said is here, tidied into sections. The audio is lightly edited: most "um"s and "uh"s cut, long pauses shortened, EQ and level evened out (10:06 recorded, 9:04 published). Then the voice was restored with Sidon (University of Tokyo) to sound closer to a studio microphone. The intro and outro are voiced by Herman Poppleberry of My Weird Prompts, over “Happy Strummin’” by Jason Shaw (music by audionautix.com, CC BY 3.0); the recording starts at 0:15. Timestamps jump to that point in the recording. All episodes: podcast.

0:15 Sick days

Hey everyone. This is Daniel Rosehill, it's 5 November 2026, and I'm in bed with our son Ezra, a 15-month-old bundle of joy and viral illness. He's down with a cold. We used to do a lot of Friday night dinners, and I bumped into a friend yesterday who's heard us say we're sick, we can't host, we're not going out to meals, every weekend. Have you just been sick permanently for the last month? Basically yes, to greater and lesser degrees. Ezra keeps coming down with these viruses, we catch them, and life goes on in its merry way.

So I'm recording this with one eye on Ezra. Last night I was doing a sprint to get these ASR models, the project I'm calling Hebrish, to a really good point of, I would say, stability.

1:22 Why the site says why I built this

I was also editing the website, of course with Claude Code. Who does anything without an AI agent now, especially make a website? Going through the home page making light-touch edits, I realised it didn't really say anything about why I built this.

That isn't because I want my branding and life story all over the site. It's because I really want this to be free for ordinary users. I've done a bit of thinking about whether there's a way to do some commercial licensing, and as the website says, it's not ready for that yet. But whether or not it ever is, for non-commercial use I'm 100% committed to this being a free resource for the English-speaking Jewish world and immigrants in Israel, because I think it's an important resource.

2:25 Recording 25 hours of sentences

So I fleshed out a few details about my bio: how I got into voice dictation, and how I became interested enough to spend 25 hours of my week recording sentences into my phone at various hours of the day and night to build up training data.

I'm creating this first and foremost for my own use. I spend hours a day dictating, and it's been about two years since I began doing it all the time. What's been very fulfilling this week is moving from the early proof-of-concept data sets to data sets of around 25 hours. That's enough to get a really good sense of how much difference a fine-tune is going to make.

The routine

Scaling the recordings literally just means using an app that Claude Code made. It's a fantastic app, and I kind of zone out and find it relaxing. It's my routine while I'm waiting for my coffee or, more honestly, my morning ADHD medicine to kick in: okay, I'm going to read 100 sentences. So if I ever sound a bit groggy, I probably am.

I've also been going out to record in different environments, to get different noise profiles and noise floors, picking up my phone or computer through the day and going at it for another half hour. The app is almost gamified, slightly addictive, which is weird for an activity that is by anyone's account quite mind-numbing. Just one more, just one more.

4:19 Biblical sentences

I added biblical sentences this week, partly for the vocabulary. Some Bible names aren't ones you'll encounter very often in modern Israel, but the Bible is such a strong part of our Jewish tradition that I really wanted its verses in there. They're the ones I love reading. Sometimes I just read those and absorb the Tanakh through the exercise, which is quite enjoyable.

5:08 One model, then incremental training

Two domain vocabularies are tricky for ASR in my case: tech vocabulary, including a lot of specialist AI words, and Hebrew. Once it was clear that neither task got worse from including the other, it made sense to bundle everything into one model.

Sabich Rolling [now simply Sabich in the apps] split off because I wanted to try my first incremental checkpoint. If you know a bit about fine-tuning: instead of going back to the raw model and doing the whole thing again, you carry on as the data set grows. Fine-tuning takes the base model and redistributes its weights by putting it through what are called epochs, or runs over the new data, and as the data set grows you ideally don't want to redo all of that on every iteration.

I called the first try the rolling release, just so I could test it for myself for a day. I didn't see any loss in quality or overfitting on the delta, so that's what I'm going to maintain.

6:34 Two models to maintain

Everything has a DOI, so I can't really rename the repos. But Rolling is the latest one, and the stable one might get a little training now and then. Two models is about the maximum I can keep training, and the second is the Nemotron streaming one. So it's a two-horse race.

• Maintained: Sabich and Shakshuka, the codename for the Nemotron streaming model. These are the ones I'm committed to retraining as the data set grows.
• Parked: Jachnun and Parakeet 100M [Sabich-Parakeet-110M]. All the early models are still there; they were all trained.

Yesterday I looked at training Shakshuka on Nemotron 3.5, the multilingual one. For technical reasons it didn't make sense, so we're staying on the slightly older version of their streaming model.

7:40 Sentences in the wild

On the data set this week, I got some sentences in the wild, including one from a friend who dropped by the apartment yesterday. Jason, thank you for your sentence. As soon as my family is over our illnesses, I have another friend I really want to read the Yiddish stuff. And I'm adding more domain vocabulary across the areas.

8:09 Where it stands

I'm enjoying this and looking forward to using it. It's really great to put things into Google Maps and get street names transcribed accurately without having to spell them. And it's really fun getting deep into fine-tuning ASR models and data set weighting.

Weighting is, I think, the more interesting challenge now, besides the most fundamental one of getting more speakers and different accents. Blending the data set becomes its own problem: avoiding internal contradictions as the scale grows.

It's still niche, still boutique, but I'm gradually telling people about it. It's free, no catches, and it works offline. Download it for Android or Windows, use it, and tell me how it works. Be honest. There's a feedback form as well. We'll keep iterating from here.

That's it from me, Daniel Rosehill, next to Ezra, who is knocked out now. This has been the Hebrish podcast.