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 to this model is by request and 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.
Meorav v2.1 (GGUF): experimental, for phones and desktops
Quantized GGUF builds of Meorav v2.1, a single-speaker text-to-speech fine-tune of Qwen3-TTS-12Hz-1.7B on one voice: Daniel Rosehill, Irish-accented English with Hebrew words mixed in ("Hebrish"). For qwentts.cpp, a C++/GGML port of Qwen3-TTS that runs on CPU (including Android) and on GPUs through Vulkan.
Experimental, public testing. These files have not been listened to against the original model. Automatic checks only (below).
Access and terms of use. Governed by the Hebrish terms of use; access is by request (button on this page, or access@hebrish.app). Non-commercial use is free; commercial use needs a commercial licence.
Files
A voice needs two files: the talker (text to speech codes) and the codec (codes to audio). Use a matching pair.
| Quant | Talker | Codec | Peak RAM while speaking | For |
|---|---|---|---|---|
| Q8_0 | qwen-talker-1.7b-customvoice-Q8_0.gguf (2.04 GB) |
qwen-tokenizer-12hz-Q8_0.gguf (291 MB) |
~5.8 GB | Desktops; phones with 12 GB+ |
| Q4_K_M | qwen-talker-1.7b-customvoice-Q4_K_M.gguf (1.18 GB) |
qwen-tokenizer-12hz-Q4_K_M.gguf (255 MB) |
~4.1 GB | Phones |
Checksums are in SHA256SUMS. The voice is the built-in speaker daniel (speaker id 3000); no
reference clip is needed.
Use
git clone --recursive https://github.com/ServeurpersoCom/qwentts.cpp && cd qwentts.cpp
./buildcpu.sh # needs OpenBLAS; ./buildvulkan.sh for a GPU; ./buildtermux.sh on Android
echo "I popped into the makolet on the way home." | \
./build/qwen-tts --model qwen-talker-1.7b-customvoice-Q4_K_M.gguf \
--codec qwen-tokenizer-12hz-Q4_K_M.gguf \
--speaker daniel --lang English -o out.wav
Output is 24 kHz mono. Write numbers as words ("twenty past six"): the model was trained that way.
How they were made
From Hebrish/Meorav at fdcb65dd (v2.1), with qwentts.cpp at 5916bf7 (2026-10-16): its convert.py
(no changes; it reads the fine-tune's speaker table and tts_model_type) to F32 GGUF, then its quantize
tool. The F32 files are not published; they can be rebuilt in about a minute.
Checks (2026-10-22, automatic only)
Two test lines, two seeds, each quant, against the original model in PyTorch (bf16):
- Every file produced intelligible speech with speaker
daniel: no silence, no runaway output, no clipping. - Durations match the original within about 10–20% (e.g. 6.8–7.8 s vs 6.8–8.2 s on the same line).
- A Whisper small.en transcript of an English line is near-exact at both quants. Hebrew words come out equally approximate from the GGUF and the original, so that is the model, not the conversion.
- Speed on a laptop CPU (Ryzen 7 5700U, no BLAS, other work running): Q4_K_M 1.9–2.8× slower than real time, Q8_0 2.2–3.4×. Not real time on that machine; phone speed not yet measured.
Not checked: that it sounds like the original by ear, and speaker similarity by embedding.
Base model and licence notices
This model is a build of Hebrish/Meorav, 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 and converted the weights for this build.
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.
- Downloads last month
- 3
4-bit
8-bit