You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access is approved automatically once you accept the terms. Use is governed by the Hebrish terms of use (https://hebrish.app/terms): non-commercial use is free; commercial use needs a commercial licence from access@hebrish.app.

Log in or Sign Up to review the conditions and access this model content.

Configuration Parsing Warning:In adapter_config.json: "peft.task_type" must be a string

Meorav v3.0 (proof of concept)

English text-to-speech that says the Hebrew words in it the Israeli way ("Hebrish"), in one synthetic voice: a generic young American male narrator, designed from attributes and cloned from nobody. Meorav is Hebrew for "mixed", as code-switching is. Released 2026-11-15.

v3 replaces Meorav v1-v2.2, which spoke in Daniel Rosehill's own voice. It is a different model on a different base: a LoRA adapter for k2-fsa/OmniVoice, not a Qwen3-TTS fine-tune. v2.2 is kept at the git tag v2.2.

Daniel's own voice now has its own repo: Hebrish/Meorav-Daniel (v1.0 = Meorav v2.2, unchanged, 2026-11-15). Meorav carries the neutral designed voice; Meorav-Daniel the faithful clone.

Five designed voices in one model: Hebrish/Meorav-Multi-Voice (v1.0, 2026-11-16). American male and female, British male and female, Australian male, picked by a reference clip; a Qwen3-TTS fine-tune continued from Meorav-Daniel, trained on the Hebrew spans of Daniel's takes only.

This is a first proof of concept: 650 training takes (60 min), not benchmarked, no listening score reported.

Listen

Generated by this adapter with no reference audio (the adapter alone is the voice). None of these sentences were picked by ear; the last two are not code-switched.

Text Audio
Hi Dana, I'll send you the final schedule machar morning, since we're still waiting on the hotel confirmation.
Since the milchama started, our whole team has been working remotely.
I finally got the letter from Bituach Leumi about my benefits.
We drove up from Haifa to the shuk in Akko, and honestly, it was a total balagan.
The town sat at the edge of the valley, where the river slowed and widened. Every morning, the first thing you heard was the bakery opening its shutters.
And here's the thing nobody tells you about benchmarks. The number on the leaderboard was measured on clean read speech, in a quiet room, by people who knew they were being recorded.
Yeah.

Demo: "The Voice from Nachlaot" (7 min 45 s)

A short story about the day this model was made, written for it and read by it: one take per sentence, no reference audio, no edits. It contains 78 Hebrew terms (61 distinct). Text: demo/the-voice-from-nachlaot.md (also in this repo under demo/).

Whisper large-v3-turbo transcribes it with a 96.5% word match to the text. The demo audio and text are CC-BY-NC-4.0 like the model, and may be redistributed with it.

How it was made

Every synthetic route to Hebrew pronunciation tried first failed by ear: OmniVoice reading pointed Hebrew script (1 of 11 terms right), ARPAbet overrides (73% right or close, no guttural, disjointed), and a phoneme TTS (Zonos) fed IPA. What worked was keeping real speech and changing only the voice:

  1. Voice. Designed with OmniVoice voice design (male, young adult, moderate pitch, american accent), picked by ear from 20 candidates, its reference clip restored with Sidon v0.1.
  2. Source speech. 650 takes (60 min) of Daniel Rosehill's recordings from Hebrish/Hebrish-TTS-Voice-Recordings: 500 code-switched takes chosen to cover his most-used Hebrew terms (945 terms, 68% of all Hebrew-term occurrences in his recordings) and 150 English takes. These carry the correct Israeli pronunciation.
  3. Voice conversion. Each take converted to the designed voice with Seed-VC v1 (zero-shot, timbre only), then enhanced with NVIDIA NemotronLabs RE-USE. Chosen in blind pairwise listening rounds over fine-tuned Seed-VC, RVC, kNN-VC and StyleStream accent conversion (which broke the Hebrew).
  4. Training. An OmniVoice LoRA (r 16, alpha 32, lr 1e-4, 1,000 steps, one L40S, about 9 minutes) on the converted audio paired with the original romanized text, with OmniVoice's default condition dropout so it speaks without a reference clip.

Use

pip install "omnivoice[lora] @ git+https://github.com/k2-fsa/OmniVoice@08be0b4" peft soundfile
import soundfile as sf, torch
from huggingface_hub import snapshot_download
from omnivoice import OmniVoice
from omnivoice.utils.lora import load_lora_adapter

base = OmniVoice.from_pretrained("k2-fsa/OmniVoice", revision="c5fdb5ccb189668d56333f77ba2629f4cd7535f4",
                                 device_map="cuda:0", dtype=torch.float16)
model = load_lora_adapter(base, snapshot_download("Hebrish/Meorav"))
audio = model.generate(text="I'll send it machar morning, after the shuk.", language="English",
                       num_step=32, guidance_scale=2.0, denoise=True, postprocess_output=True)[0]
sf.write("out.wav", audio, model.sampling_rate)

The PyPI OmniVoice wheel (0.2.1) has no LoRA support; install from git as above.

Limitations

  • Proof of concept, trained on 60 minutes. Hebrew terms outside the 945 covered are untested.
  • Some of the source speaker's accent remains. Voice conversion changes the voice, not the accent; accent conversion was tested and broke the Hebrew words.
  • Not benchmarked. Automatic check on the samples above: Whisper large-v3-turbo transcribes every word, and speaker similarity to the designed voice's reference is 0.64-0.83 (ECAPA, clips over 2 s).

Licence and attribution

  • Non-commercial: CC-BY-NC-4.0 plus the Hebrish terms of use. Commercial use needs a commercial licence.
  • Base model: k2-fsa/OmniVoice. Its code is Apache-2.0; its pretrained weights are CC-BY-NC "due to constraints from its training data (e.g., Emilia)". This adapter inherits that non-commercial restriction.
  • Training data pipeline: Seed-VC (GPL-3.0 code, used only to convert data), NVIDIA NemotronLabs RE-USE (OpenMDW-1.1), Sidon v0.1 (MIT).
  • The voice is synthetic. It was designed from attributes and imitates no real person. The source recordings are Daniel Rosehill's, published with his consent in the dataset above.

Listen: what Hebrish is and how it's made

A spoken introduction to the project by the person who makes it: why Hebrish exists, why stock speech-to-text gets Hebrew words wrong, how the fine-tunes are trained, the model lines (Sabich, Jachnun, Shakshuka, Meorav), how the training data is recorded, and the plans for the apps.

Listen hebrish.app/updates: player, chapters, edited transcript, MP3 download
Author Daniel Rosehill
Recorded 2026-10-22
Runtime 1:07:03

Chapters (each opens the player at that point): 0:00 Why this voice note · 1:36 Coming to speech to text · 6:36 Why Hebrew words go wrong · 9:23 How the fine-tune works · 12:27 Why Parakeet, and the model plan · 17:44 Disfluencies, shekels, paragraphs · 25:40 Tech vocabulary and Meorav · 29:07 The kitchen sink: Sabich and Jachnun · 33:35 Licensing, accents, noisy recordings · 39:10 One recorder for ASR and TTS · 41:04 Desktop shortcuts and the macro pad · 43:34 Android: Bluetooth and one button · 45:35 Streaming and Shakshuka · 50:28 Long-form transcription · 56:00 Output modes · 63:01 Report a word, and more voices

Downloads last month
52
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Hebrish/Meorav

Finetuned
Qwen/Qwen3-0.6B
Finetuned
k2-fsa/OmniVoice
Adapter
(60)
this model
Quantizations
1 model

Dataset used to train Hebrish/Meorav

Spaces using Hebrish/Meorav 5

Collections including Hebrish/Meorav