Meorav-Multi-Voice / README.md
danielrosehill's picture
Card: voice names (Josh, Rivka, Nigel, Fiona, Bazza) and avatars, proof-of-concept badge
210e728 verified
|
Raw History Blame Contribute Delete
6.46 kB
metadata
license: cc-by-nc-4.0
base_model: Qwen/Qwen3-TTS-12Hz-1.7B-Base
base_model_relation: finetune
language:
  - en
  - he
pipeline_tag: text-to-speech
tags:
  - tts
  - qwen3-tts
  - voice-clone
  - meorav
  - code-switching
  - hebrish
  - multi-speaker
extra_gated_prompt: >-
  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.
extra_gated_fields:
  Name: text
  Organisation: text
  Intended use: text
  I agree to use this model under the Hebrish terms of use (non-commercial unless I hold a commercial licence): checkbox
  I agree to receive Hebrish model updates by email (at most once a month; unsubscribe any time): checkbox

Meorav-Multi-Voice

Status: proof of concept

Proof of concept. v1.0 was not trained on enough data, so the Hebrew code-switching did not transfer as well as hoped: gutturals such as the "ch" in chag and achi are not always produced. Listen before you rely on it: Meorav-Samples. More: https://hebrish.app/models/meorav-multi-voice.

English text-to-speech with Hebrew words mixed in ("Hebrish"), in five designed voices from one model. It is a full fine-tune of Qwen3-TTS-12Hz-1.7B-Base, continued from Meorav-Daniel. The voice is picked by a short reference clip (x-vector conditioning); the clips for the five voices are in voices/.

None of the five voices is a real person: each was designed from a text description with OmniVoice ("<gender>, young adult, moderate pitch, <accent> accent"), then fixed to one seed.

Avatar Name Voice Reference clip
Josh Josh American male (A3) voices/american-male.wav
Rivka Rivka American female voices/american-female.wav
Nigel Nigel British male voices/british-male.wav
Fiona Fiona British female voices/british-female.wav
Bazza Bazza Australian male voices/australian-male.wav

Pass the reference clip for the voice you want (ref_audio, see Usage). The names and avatars are illustrations, not real people; copies of the avatars are in figures/voices/ (the card shows them from the public Meorav-Samples Space, since files in this gated repo need access).

Access and terms of use. Governed by the Hebrish terms of use; access is approved automatically once you accept them (button on this page, or access@hebrish.app). Non-commercial use is free; commercial use needs a commercial licence.

Related models

Model Voice
Meorav-Daniel Daniel Rosehill's own voice (Irish-accented English)
Meorav One designed American male voice (OmniVoice LoRA, v3.0)
Meorav-Multi-Voice (this repo) Five designed voices, one model

Training (run mv1, 2026-11-16)

  • Init: Meorav-Daniel weights with the base model's speaker encoder restored, so new voices can be conditioned by x-vector.
  • Data: 2,147 rows. 461 of Daniel Rosehill's code-switching takes, counted twice (922 rows), trained only on their Hebrew-word spans (the English parts are masked out of the loss), so the voices learn how the Hebrew words are said without taking on his voice. Plus 245 English lines per designed voice (about 27–31 min each), generated with OmniVoice, trained in full. Every row is conditioned on its own speaker's x-vector.
  • Recipe: 3 epochs, lr 1e-5, batch 4 × grad accum 4 (effective 16), one A100-80GB on Modal. This is the epoch-3 checkpoint.
  • Training code: Qwen3-TTS fine-tuning at commit 022e286b, with a span-masked SFT script.

Usage

import torch, soundfile as sf
from huggingface_hub import hf_hub_download
from qwen_tts import Qwen3TTSModel

repo = "Hebrish/Meorav-Multi-Voice"
tts = Qwen3TTSModel.from_pretrained(repo, device_map="cuda:0", dtype=torch.bfloat16)
ref = hf_hub_download(repo, "voices/british-female.wav")  # Fiona
wavs, sr = tts.generate_voice_clone(text="Yalla, let's go, we're already late for the brit mila.",
                                    language="English", ref_audio=ref, x_vector_only_mode=True)
sf.write("out.wav", wavs[0], sr)

Requires qwen-tts. The repo is gated: log in with a token that has access. Write Hebrew words in Latin letters, the way an English speaker in Israel would spell them.

Limitations

  • Not benchmarked: no listening score or metric is reported. The renders were checked by ear only.
  • Gutturals (ח, the "ch" in chag, achi) are not always produced.
  • Training clips were at most 15 s; generate long text sentence by sentence.
  • Other voices can be tried with any reference clip, but only the five above were trained.

Licence

CC BY-NC 4.0. The base model is Apache 2.0.

Base model and licence notices

This model is a fine-tune of Qwen/Qwen3-TTS-12Hz-1.7B-Base, Copyright 2026 Alibaba Cloud, licensed under Apache-2.0 (text in LICENSE-Qwen-Apache-2.0). Hebrish modified the base model by fine-tuning it.

The licence stated at the top of this card covers Hebrish's contribution; the base model's licence keeps applying to the base model, and nothing here restricts what that licence allows for it. Terms of use: https://hebrish.app/terms.

Hebrish is not affiliated with or endorsed by Alibaba Cloud.