Models · Speech synthesis · Qwen3-TTS 12Hz 1.7B Base
Meorav
Text-to-speech for code-switched English: Irish-accented English that pronounces the Hebrew words in it correctly. A full fine-tune of Qwen3-TTS, and the Hebrish line that goes the other way, from text to speech. The current version is Meorav v2.2, released 2026-11-09 (release post, listen to samples, Android app), trained on about 5.5 hours of the speaker's code-switching takes. How it's made lists the recipe and every model used.
What it is
Give it a line of English with Hebrew words in Latin letters, and it reads the line aloud, with the Hebrew said the Israeli way rather than spelled out as English:
Yalla, tachles: Bituach Leumi wants the form by Thursday.
Meorav is a full fine-tune (every weight updated) of Qwen3-TTS-12Hz-1.7B-Base. Qwen3-TTS reads raw text through its tokenizer, with no phonemiser in between, and a full fine-tune lets the model move off the American accent it starts with. Two earlier attempts at this voice, an Orpheus fine-tune and a zero-shot clone, came out American: a zero-shot clone copies the timbre but takes its prosody from the base model.
Meorav is Hebrew for "mixed", which is what code-switching is, and meorav yerushalmi is the Jerusalem mixed grill.
The other route to the same goal is the deterministic English-Hebrew Code-Switching TTS Pipeline (romanized Hebrew, then Hebrew script with niqqud, then a Hebrew voice for each Hebrew span; see the token classifier); Meorav is one model that learned the Hebrew words from recordings instead.
Samples
Ten sentences, shortest first, each as a Before and After pair plus two more takes from Meorav, as they came out of the models; none was picked or edited. Before is the untuned base model, Qwen3-TTS-12Hz-1.7B-Base, cloning Daniel's voice zero-shot from one reference recording (the same reference take training used). After is Meorav v2.2 at the default sampling temperature (0.9). The last two players are the same Meorav weights at temperature 1.1 and 1.3: a higher temperature lets the model pick less likely prosody, which can sound livelier and less flat, at some risk of odd stresses. Short texts are generated in one pass; longer ones sentence by sentence, then joined.
Yalla, let's go, we're already late for the brit mila.
We're doing a quiet shabbat this week, just challah, some chamin and an early night.
I spent the whole morning at Misrad Hapnim, and I still don't have the teudat zehut.
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.
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.
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.
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.
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.
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.
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.
For comparison with Daniel's real recordings and with the base model, see Hebrish/Meorav-Samples on Hugging Face.
Status
Meorav v2.2 (2026-11-09, run meorav-v22-reuse, release notes) is the current version. It is published as Hebrish/Meorav (revision 936eea1a), gated (access on request, like the other Hebrish models), with a GGUF build in Hebrish/Meorav-GGUF and on the Hebrish CDN at cdn.hebrish.app/models/meorav/. The training settings are v2.1's; the data is new: 3,995 screened code-switching takes, about 5.5 hours (v2.1 had 1,103 takes, about 88 minutes), enhanced with NVIDIA NemotronLabs RE-USE instead of DeepFilterNet 3. How it's made has the details. Daniel listened to its samples before it was released.
Meorav v2.1 (2026-10-22, run meorav-v2-df, part of Release v2) was the previous version. It is published as Hebrish/Meorav, gated (access on request, like the other Hebrish models). The recipe is the same as v2.0's (a full fine-tune of Qwen3-TTS-12Hz-1.7B-Base, learning rate 7e-6, 3 epochs), but the training audio was first cleaned with DeepFilterNet 3, a noise suppressor: 1,103 takes, 88.2 minutes. Daniel chose it over v2.0 by listening to both. On 7 test sentences, the background noise below 1 kHz is about 5 dB lower than v2.0's, with the same voice bandwidth; nothing else has been measured.
Meorav v2.0 (2026-10-22, run meorav-v2) was the first release in the same repository, replaced by v2.1 the same day: 1,103 quiet-room takes, 87.8 minutes, 3 epochs, final epoch kept. It stays available at revision 600a2820. Meorav versions follow the release number, so there is no Meorav v1.0.
Meorav PoC-1 POC was the first training run (first-test, 2026-10-22): 657 read takes, 50 minutes after trimming, from 48 kHz / 24-bit recordings resampled to 24 kHz for training. Learning rate 7e-6, 3 epochs, batch 2 with gradient accumulation 4, on one L40S on Modal. It was published, open, as Hebrish/TTS-V1 (a name that predates Meorav); that repository has since been removed from Hugging Face, so PoC-1 is no longer available. The current versions are in Hebrish/Meorav.
No benchmark exists yet. None of the versions has been evaluated with a metric; v2.1 and v2.2 were chosen by listening.
How it's made
Meorav v2.2 is the first version trained on this recipe. Every step and every model used:
- Screening. Candidate takes come from Daniel's own recordings, including the takes he records for Hebrish's speech recognition data. Each take gets sound-event tags from PANNs (pretrained audio neural networks trained on AudioSet) and quality scores from Microsoft's DNSMOS. Only code-switching takes (English with Hebrew words) are kept: no tech vocabulary and no noisy rooms. Takes containing digits were dropped, and takes over 15 seconds were left out of training.
- Enhancement. Daniel's takes are recorded at home: real rooms, a crying baby now and then, a phone or desktop microphone. Before training, every kept take passes through NVIDIA NemotronLabs RE-USE (
nvidia/NVIDIA-NemotronLabs-RE-USE, OpenMDW-1.1 licence), a speech-restoration model that removes background noise and room sound and keeps the voice. It runs as theaudio-cpp/RE-USE-GGUFbuild through audio.cpp. This step matters because a text-to-speech model learns everything in its training audio, the noise included: cleaner input means cleaner output. RE-USE replaced DeepFilterNet 3, which cleaned the audio for v2.1. Sidon v0.1 was also tried: both Sidon and RE-USE removed a crying baby from a take completely, without the robotic tinge older noise suppressors leave. Daniel gave RE-USE a slight edge by ear, and a spectrogram comparison decided it: RE-USE keeps the speaker's own detail above 4 kHz, where Sidon re-synthesises it. - Mastering and training. The enhanced takes go through a podcast-style mastering chain (EQ, de-esser, compressor, loudness to -20 LUFS), then train a full fine-tune (SFT, every weight updated) of
Qwen/Qwen3-TTS-12Hz-1.7B-Base: 3 epochs, learning rate 7e-6, the same settings as v2.1. - Data. 3,995 takes, about 5.5 hours of audio. v2.1 had 1,103 takes, about 88 minutes.
- Checks. Hebrish's own speech recognition model, Sabich, transcribes enhanced takes to check that the Hebrew terms survive enhancement. Daniel's listening notes: the crying baby is gone and there is no robotic tinge.
Audio enhancement by NVIDIA NemotronLabs RE-USE. Every training take for Meorav v2.2 was cleaned with NVIDIA NemotronLabs RE-USE (OpenMDW-1.1 licence), run through audio.cpp from the audio-cpp/RE-USE-GGUF build.
Models used: Qwen3-TTS-12Hz-1.7B-Base (base model), NVIDIA NemotronLabs RE-USE via audio-cpp/RE-USE-GGUF (enhancement), PANNs and DNSMOS (screening), and Sabich (term check). The result is in Hebrish/Meorav and Hebrish/Meorav-GGUF.
Recordings
Meorav is trained on one speaker, Daniel Rosehill, and speaks in his voice. The takes are in Hebrish/Hebrish-TTS-Voice-Recordings (CC BY-NC 4.0): mono WAV at 48 kHz / 24-bit, with no browser processing, read from prompt sets that cover English phonetics, everyday conversation, English sentences carrying Hebrew words and Israeli names and places, numbers, tech vocabulary and free speech.
They are recorded in a dedicated quiet-room voice-clone mode of the phone capture app that also collects Hebrish's speech recognition data. The two need opposite conditions: speech recognition training uses every take, including deliberately noisy ones (an air conditioner, people talking), while a voice clone needs a quiet room. Up to v2.1, only voice-clone takes that passed the level checks trained Meorav. From v2.2, quiet code-switching takes from the speech recognition recordings are screened in too (see How it's made).
Use it
On an Android phone: Meorav (beta)
The Meorav app runs the model on the phone with qwentts.cpp: type a line, and it reads it aloud in Meorav's voice. On first use it downloads the voice (Meorav v2.2) from the Hebrish CDN; after that it works offline. It is a beta, and Android only.
Download Meorav (beta) 0.1.0 for Android APK, 2.99 MB, plus the voice on first use. Release page
Hardware for generating speech
Approximate figures. "Measured" means timed on that hardware; "estimate" means not measured yet.
| Where it runs | Model files | Speed and memory | |
|---|---|---|---|
| Download size, on-device GGUF (qwentts.cpp) | talker + codec: Q4_K_M 1.18 GB + 255 MB, or Q8_0 2.04 GB + 291 MB | — | File sizes |
| Laptop CPU | GGUF Q4_K_M | 1.9 to 2.8 times slower than real time; about 4.1 GB peak RAM | Measured |
| Desktop GPU through Vulkan (AMD Radeon RX 7700 XT) | GGUF Q8_0 | 5.8 s of audio in 6.7 s, including loading the model: about real time | Measured |
| Phone | GGUF (in the Meorav app) | Slower than real time; a recent phone with 6 GB of RAM or more is the sensible minimum | Estimate |
Full precision, PyTorch (qwen-tts) | 1.7B parameters: 3.8 GB bf16 checkpoint + 0.7 GB codec | An NVIDIA GPU with about 8 GB of VRAM or more; faster than real time on data-centre GPUs | Estimate |
In Python
import torch, soundfile as sf
from qwen_tts import Qwen3TTSModel
tts = Qwen3TTSModel.from_pretrained("Hebrish/Meorav", device_map="cuda:0", dtype=torch.bfloat16)
wavs, sr = tts.generate_custom_voice(text="Yalla, I'll ring you after the match.", speaker="daniel", language="English")
sf.write("out.wav", wavs[0], sr)
Needs qwen-tts (tested with 0.1.1) and a GPU. Numbers are spelled out in the training text, so write them out in prompts too ("four hundred and twelve"). From the model card. Hebrish/Meorav is gated: request access first.
The training pipeline is public: Hebrish-ASR/Meorav-TTS has the data preparation, the Modal training stages, the run log, and a patched copy of Qwen3-TTS's fine-tuning script with fixes useful to anyone fine-tuning it. Its README explains how to adapt it to another voice.
At a glance
| Base model | Qwen/Qwen3-TTS-12Hz-1.7B-Base (Qwen) |
| Current version | Meorav v2.2, released 2026-11-09 |
| Parameters | 1.9B (1,916,676,352 in model.safetensors, Hub safetensors count; plus a 171M speech tokenizer) |
| Evaluation | No benchmark yet; v2.2 released after Daniel's listening check Status on this page · Training pipeline and run log |
| Access | Gated, by approval: you request access on Hugging Face and a person reviews each request. Request access on Hugging Face |
| Download | every format and quant (Hugging Face) |
| Collection | Hebrish/hebrish-english-hebrew-code-switching on Hugging Face |
| Licence | CC-BY-NC-4.0 · use governed by the terms of use |
| Used by | Your own tools (qwen-tts); the Hebrish apps do not offer text-to-speech |
| DOIs | none yet |
Repository
| Repository | Format and files | Version | For | DOI |
|---|---|---|---|---|
Hebrish/Meorav main | Qwen3-TTS (safetensors) model 3.83 GB (bf16), speech tokenizer, training config · needs approval · files | v2.2 | qwen-tts, voice "daniel" | none yet |
Hebrish/Meorav-GGUF | GGUF talker and tokenizer GGUF files, SHA256SUMS; the v2.2 GGUF is also on cdn.hebrish.app/models/meorav/ · needs approval · files | see the model card | llama.cpp-style runtimes | none yet |
Hugging Face issues a DOI per repository, so each format can be cited on its own. Repositories, sizes and DOIs checked on the Hub on 2026-11-09.
Versions
| Version | Released | Run | Where | Notes |
|---|---|---|---|---|
| Meorav v2.2 current | 2026-11-09 | meorav-v22-reuse | Meorav @ 936eea1a | New data recipe: screened code-switching takes, enhanced with NVIDIA NemotronLabs RE-USE instead of DeepFilterNet 3 (3,995 takes, about 5.5 h); same training settings as v2.1; released after Daniel's listening check; not evaluated yet |
| Meorav v2.1 | 2026-10-22 | meorav-v2-df | Meorav @ fdcb65dd | Release v2, same-day update: same recipe as v2.0, takes cleaned with DeepFilterNet 3 first (1,103 takes, 88.2 min); chosen by listening; not evaluated yet |
| Meorav v2.0 | 2026-10-22 | meorav-v2 | Meorav @ 600a2820 | Release v2: 1,103 quiet-room takes, 87.8 min; 3 epochs, final epoch kept; replaced by v2.1 the same day |
| Meorav PoC-1 POC published as TTS-V1 | 2026-10-22 | first-test | no longer published | 657 read takes, 50 min after trimming; first test, not benchmarked; no longer on Hugging Face (Hebrish/TTS-V1 removed; noticed 2026-10-27) |
Cite
Meorav has no DOI yet; cite the repository, Hebrish/Meorav.
@misc{rosehill_meorav_2026,
author = {Daniel Rosehill},
title = {Meorav},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/Hebrish/Meorav}
}
Licence and access
Request access on Hugging Face first. The main Meorav repository is gated: open it, sign in, and fill in the short request form there. Requests are reviewed by a person.
For non-commercial use, the attestation form covers what is asked.
No answer after a few days? Request access here instead, or write to access@hebrish.app. Commercial use, including telephony, needs a commercial licence.