Model · text-to-speech
Meorav: speech synthesis for English–Hebrew code-switching
A text-to-speech model for English with Hebrew words dropped in, written in Latin letters, the way English speakers living in Israel talk. In my voice, Irish accent included.
Listen: ten sentences, four ways
Every sentence below is English with Hebrew words dropped in, the way English speakers living in Israel actually talk. Listen to the Hebrew words: Misrad Hapnim, teudat zehut, acharei hachagim. Most English text-to-speech voices read those like a tourist with a phrasebook (wrong stress, a ch that comes out as a k, Hebrew read as if it were English spelling). Meorav says them the way an English speaker who lives here says them mid-sentence: as Hebrew words, not English spelling, without a pause or a change of voice. It doesn't always get them right; listen and judge for yourself.
And yes, it has an Irish accent. It is trained on one voice, mine, and I am Irish and live in Jerusalem. I did not try to remove the accent, and this is no claim of native Hebrew either: I am not a native speaker, and Meorav does not speak Hebrew like one. What it demonstrates is the code-switching case.
At its default settings Meorav reads a little flat, like someone reading a sentence off a card (which is more or less what it was trained on). Sampling at a higher temperature lets it pick livelier phrasing, so every sentence is also here at temperature 1.1 and 1.3 to compare.
- Meorav on hebrish.appModel page: how it is made, example code, hardware notes
- Launch post on hebrish.appSpeech synthesis for English-Hebrew code-switching
- Weights on Hugging FaceGated, by request
- GGUF buildsFor running it on a device
- Samples SpaceThe same samples on Hugging Face
- Beta Android appRuns Meorav on the phone
Before, after and two temperatures
Before is the untuned base model, Qwen3-TTS-12Hz-1.7B-Base, copying my voice from a single recording. After is Meorav v2.2 at its default sampling. The last two columns are the same Meorav weights and text at a higher temperature. None of the samples was picked or edited. Passages marked "sentence by sentence" were generated one sentence at a time and joined.
| Text | Before · base model | After · Meorav | Meorav · temperature 1.1 | Meorav · temperature 1.3 |
|---|---|---|---|---|
| Yalla, let's go, we're already late for the brit mila. | 2.7 s | 3.5 s | 2.9 s | 3.3 s |
| I spent the whole morning at Misrad Hapnim, and I still don't have the teudat zehut. | 4.3 s | 4.6 s | 4.2 s | 4.8 s |
| We're doing a quiet shabbat this week, just challah, some chamin and an early night. | 5.6 s | 4.0 s | 4.2 s | 4.1 s |
| On Friday morning I went down to the shuk for the usual run. The guy at the bourekas stall remembered me, gave me an extra one, and told me to say shabbat shalom to the kids. | 9.0 s | 8.7 s | 9.4 s | 9.0 s |
| The arnona bill came in twice this month, so I rang the iriya. After forty minutes on hold they told me, beseder, it's a mistake, just ignore the second one. | 8.9 s | 8.9 s | 8.0 s | 8.0 s |
| Everyone here says everything will happen acharei hachagim. I used to find it frustrating, but honestly, after a few years you start saying it yourself. | 8.0 s | 8.6 s | 6.9 s | 6.9 s |
| When I started ulpan, the morah told us not to worry about the grammar yet. Just listen, she said, and the words will come. She was right about most of it, although I still mix up my verb patterns when I'm tired. These days I can argue with the makolet guy about the price of cottage cheese, which feels like real progress. (sentence by sentence) | 17.7 s | 18.3 s | 18.7 s | 18.1 s |
| Last chol hamoed we did a tiyul up in the Galil with the whole mishpacha. We stopped for shakshuka at a little place near the Kinneret, then walked along the Jordan for an hour. My father-in-law kept telling everyone it was nothing compared to the old days, before the road was paved. By the time we got back to Yerushalayim it was nearly midnight, and the kids were fast asleep in the back. (sentence by sentence) | 21.1 s | 22.2 s | 22.4 s | 21.3 s |
| Our shul does a kiddush after davening every shabbat, and it has become the real centre of the week. You hear about every simcha, every new oleh, and who is looking for a chavruta. The gabbai keeps a list of who is bringing what, and somehow there is always too much kugel. If you are new to the neighbourhood, honestly, just turn up, nobody will let you stand on your own for long. (sentence by sentence) | 21.4 s | 20.3 s | 21.1 s | 23.8 s |
| When we made aliyah, the first few months were a blur of paperwork. There was the misrad haklita, the kupat cholim, Bituach Leumi, and a bank that wanted three copies of everything. Our neighbour, a sabra who had seen it all before, sat us down with a cup of coffee and said, savlanut, it all gets done eventually. She was right. By our first Pesach in the country we had a lease, a car, and a teudat oleh that was already falling apart from use. Looking back now, the bureaucracy is not what I remember. What I remember is the people who helped, without being asked, and the way the whole street seemed to come out for Yom Haatzmaut. (sentence by sentence) | 39.1 s | 35.4 s | 34.7 s | 36.3 s |
About the model
Meorav is the text-to-speech line of Hebrish, my English–Hebrew code-switching speech project. Give it a sentence of English with Hebrew words written in Latin letters and it reads the whole thing aloud, Hebrew words included, in one voice.
v2.2, released on 2026-11-09, is the first version shared as something to use rather than a proof of concept. It is a full fine-tune of Qwen3-TTS-12Hz-1.7B-Base on 3,995 takes of my voice, about 5.5 hours, for 3 epochs.
The weights are on Hugging Face at Hebrish/Meorav, by request, under CC BY-NC 4.0 (free for non-commercial use with attribution), with GGUF builds for running it on a device and a beta Android app.
There is no benchmark for Meorav. It was released after listening to it, and the samples above are there so you can judge it yourself, accent and all.
Training
- v2.2 (2026-11-09): full fine-tune of Qwen3-TTS-12Hz-1.7B-Base (every weight updated) for 3 epochs on 3,995 takes, about 5.5 hours, all my voice.
- 1,103 takes were recorded for the voice model; the other 2,924 were screened from my speech recognition recordings.
- Every take was cleaned with NVIDIA NemotronLabs RE-USE, then a podcast-style voice chain.
Training data
3,995 takes of my voice, about 5.5 hours: 1,103 recorded specifically for the voice model and 2,924 screened from the recordings I made for the Hebrish speech recognition models.
Checkpoints and roadmap
Where the model is now and the checkpoints still to come. Planned stages are the current plan, not promises.
- v2.2 Done 9 November 2026
First release meant for use: 3,995 takes, about 5.5 hours, cleaned with RE-USE.
Limitations
- One voice, with an Irish accent. It is not a neutral English voice, and its Hebrew is a non-native speaker's.
- It does not always get the Hebrew words right.
- At default sampling it reads a little flat; compare the temperature 1.1 and 1.3 samples.
- No benchmark.
- Non-commercial licence (CC BY-NC 4.0), and the weights are gated.
Related
- Meorav on hebrish.appModel page
- Write-up (blog)Meorav: speech synthesis for English-Hebrew code-switching
- Launch post on hebrish.appSpeech synthesis for English-Hebrew code-switching
- Me and my voice cloneA ten-minute conversation with an earlier Meorav