Hebrish
Technical

Meorav-Multi-Voice PoC

Text-to-speech in five designed voices from one model, for English with Hebrew words in it. v1.0 is a proof of concept.

Text-to-speech in five designed voices from one model, picked by a reference clip: Josh and Rivka (American), Nigel and Fiona (British) and Bazza (Australian); none is a real person. A full fine-tune of Qwen3-TTS continued from Meorav-Daniel; Daniel's takes were trained only on their Hebrew-word spans, so the voices learn the Hebrew words without his voice. v1.0 (2026-11-16) is a proof of concept: it was not trained on enough data for the Hebrew to carry over as well as hoped; not benchmarked.

Status

Meorav-Multi-Voice v1.0 (2026-11-16, run mv1) is one model with five designed voices, picked by a short reference clip kept in the repository's voices/ folder. None is a real person: each was designed with OmniVoice from a text description. It is not benchmarked and has no listening score; Daniel listened to the renders and said they sound great, and that is the only check so far.

  • Josh, American male. Designed voice A3, the same voice as Meorav v3.0. Reference clip voices/american-male.wav.
  • Rivka, American female. Reference clip voices/american-female.wav.
  • Nigel, British male. Reference clip voices/british-male.wav.
  • Fiona, British female. Reference clip voices/british-female.wav.
  • Bazza, Australian male. Reference clip voices/australian-male.wav.

The names come from the audio drama below. The avatars are illustrations, not likenesses: there is no real person behind any of the voices.

It is a full fine-tune of Qwen3-TTS-12Hz-1.7B-Base, continued from Meorav-Daniel. Daniel's 461 code-switching takes were trained only on their Hebrew-word spans, so the voices learn how the Hebrew words are said without taking on his voice, alongside 245 OmniVoice English lines per designed voice. Three epochs on one A100-80GB on Modal. The other lines are Meorav (a single designed American voice, as an OmniVoice adapter) and Meorav-Daniel (Daniel's own voice).

Proof of concept. v1.0 PoC was not trained on enough data, so the Hebrew code-switching did not transfer to the five voices as well as hoped. The Hebrew words came from about nine minutes of Daniel's speech (the Hebrew-word spans of 461 takes), against more than two hours of synthetic English in the new voices, and the gutturals in particular are not always there. How it was trained, what did not carry over and what would fix it: How Meorav-Multi-Voice was trained.

Limitations. The gutturals (the "ch" in chag, achi) are not always produced. Training clips were at most 15 seconds, so generate long text sentence by sentence. Only the five voices were trained, though other reference clips can be tried. Write Hebrew words in Latin letters, the way an English speaker in Israel would spell them.

Base modelQwen/Qwen3-TTS-12Hz-1.7B-Base (Qwen)
Current versionMeorav-Multi-Voice v1.0, released 2026-11-16
Parameters1.93B (1,928,677,440 in model.safetensors, bf16, safetensors header count) plus the 170,557,441-parameter speech tokenizer
EvaluationNo benchmark and no listening score; checked by ear only; v1.0 is a proof of concept
Status on this page · Training pipeline and run log
AccessGated, automatic: accept the terms on Hugging Face and access is approved instantly. Accept terms and download on Hugging Face
Downloadevery format and quant (Hugging Face)
CollectionHebrish/hebrish-english-hebrew-code-switching on Hugging Face
LicenceCC-BY-NC-4.0 · use governed by the terms of use
Used byYour own tools (qwen-tts); the Hebrish apps do not offer text-to-speech
DOIsnone yet

Samples

Ten lines in each of the five designed voices, one take per line (seed 0, final checkpoint). None of the voices is a real person. Checked by ear only; not benchmarked.

Audio drama: The Teudat Zehut

The Teudat Zehut: a comedy in four queues, 8 min 31 s, first render. Daniel narrates in his own voice and the five designed voices play the rest:

  • Daniel, a new oleh, narrates in his own voice (Meorav-Daniel).
  • Bubbe Rivka, the American grandmother (Rivka, American female).
  • Nigel, the Misrad HaPnim clerk from Manchester (Nigel, British male).
  • Fiona at the bank (Fiona, British female).
  • Bazza from Perth (Bazza, Australian male).
  • Josh (American male) is the phone menus and the ticket machine.

Josh: American male

TextAudio
I parked the car near the water tower around thirty minutes after four.
Honestly, I think the weather this morning was far better than yesterday's.
We're going to start the meeting early, so bring your laptop and a charger.
The book about the history of the north shore is sitting on the third shelf.
Could you turn the heating down a little? It's getting warm in here.
I have to go to Bituach Leumi tomorrow to sort out the forms.
Pick up some bamba and a few pitot from the makolet on the way home.
My achi called to say the chag starts earlier this year.
We're meeting at the shuk in Machane Yehuda before Shabbat.
The teudat zehut renewal at Misrad HaPnim took almost two hours.

Rivka: American female

TextAudio
I parked the car near the water tower around thirty minutes after four.
Honestly, I think the weather this morning was far better than yesterday's.
We're going to start the meeting early, so bring your laptop and a charger.
The book about the history of the north shore is sitting on the third shelf.
Could you turn the heating down a little? It's getting warm in here.
I have to go to Bituach Leumi tomorrow to sort out the forms.
Pick up some bamba and a few pitot from the makolet on the way home.
My achi called to say the chag starts earlier this year.
We're meeting at the shuk in Machane Yehuda before Shabbat.
The teudat zehut renewal at Misrad HaPnim took almost two hours.

Nigel: British male

TextAudio
I parked the car near the water tower around thirty minutes after four.
Honestly, I think the weather this morning was far better than yesterday's.
We're going to start the meeting early, so bring your laptop and a charger.
The book about the history of the north shore is sitting on the third shelf.
Could you turn the heating down a little? It's getting warm in here.
I have to go to Bituach Leumi tomorrow to sort out the forms.
Pick up some bamba and a few pitot from the makolet on the way home.
My achi called to say the chag starts earlier this year.
We're meeting at the shuk in Machane Yehuda before Shabbat.
The teudat zehut renewal at Misrad HaPnim took almost two hours.

Fiona: British female

TextAudio
I parked the car near the water tower around thirty minutes after four.
Honestly, I think the weather this morning was far better than yesterday's.
We're going to start the meeting early, so bring your laptop and a charger.
The book about the history of the north shore is sitting on the third shelf.
Could you turn the heating down a little? It's getting warm in here.
I have to go to Bituach Leumi tomorrow to sort out the forms.
Pick up some bamba and a few pitot from the makolet on the way home.
My achi called to say the chag starts earlier this year.
We're meeting at the shuk in Machane Yehuda before Shabbat.
The teudat zehut renewal at Misrad HaPnim took almost two hours.

Bazza: Australian male

TextAudio
I parked the car near the water tower around thirty minutes after four.
Honestly, I think the weather this morning was far better than yesterday's.
We're going to start the meeting early, so bring your laptop and a charger.
The book about the history of the north shore is sitting on the third shelf.
Could you turn the heating down a little? It's getting warm in here.
I have to go to Bituach Leumi tomorrow to sort out the forms.
Pick up some bamba and a few pitot from the makolet on the way home.
My achi called to say the chag starts earlier this year.
We're meeting at the shuk in Machane Yehuda before Shabbat.
The teudat zehut renewal at Misrad HaPnim took almost two hours.

All demo audio is also in the public dataset Hebrish/Demos on Hugging Face.

How to try it

The model is on Hugging Face as Hebrish/Meorav-Multi-Voice: accept the terms and access is approved instantly. Run it yourself with qwen-tts; the model card has a short Python example. Pick a voice by passing its reference clip from voices/ with x_vector_only_mode=True. Write Hebrew words in Latin letters, the way an English speaker in Israel would spell them, and generate long text sentence by sentence. The Hebrish apps do not offer text-to-speech.

To try the voices without writing code, use Meorav desktop 0.2.0, a desktop tester for Windows and Linux with a picker for the five voices. It runs on the CPU (about 2 to 3 seconds of work per second of speech) and downloads about 1.5 GB on first use. Linux: sudo apt install meorav on a machine with the Hebrish repository, or the .deb from the release. Windows: unzip Meorav-0.2.0-windows-x64.zip and run Meorav.bat; it is unsigned, so Windows may warn before opening it.

Formats

RepositoryFormat and filesVersionForDOI
Hebrish/Meorav-Multi-Voice mainQwen3-TTS full fine-tune (safetensors)
model weights, tokenizer and config, five reference clips in voices/ · accept terms · files
v1.0qwen-tts (Qwen3TTSModel), x-vector reference clip from voices/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 2026-11-16.

Versions

VersionReleasedRunWhereNotes
Meorav-Multi-Voice v1.0 POC current2026-11-16mv1Meorav-Multi-Voice @ v1.0Full fine-tune of Qwen3-TTS-12Hz-1.7B-Base continued from Meorav-Daniel, 3 epochs on one A100-80GB (Modal), epoch-3 checkpoint. 2,147 rows: 461 of Daniel's code-switching takes counted twice, trained only on their Hebrew-word spans, plus 245 OmniVoice English lines per designed voice; every row conditioned on its own speaker's x-vector. Checked by ear only; not benchmarked. Gutturals (ch in chag, achi) are not always produced.

Access

Accept the terms on Hugging Face. The main Meorav-Multi-Voice repository is gated automatically: open it, sign in, accept the terms, and access is approved instantly.

Accept terms and download on Hugging Face

Cite

Meorav-Multi-Voice has no DOI yet; cite the repository, Hebrish/Meorav-Multi-Voice.

@misc{rosehill_meorav_multi_voice_2026,
  author    = {Daniel Rosehill},
  title     = {Meorav-Multi-Voice},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/Hebrish/Meorav-Multi-Voice}
}