Ganj voice (آوای گنج)

A Persian text-to-speech voice for classical poetry and prose. It speaks in "nobody's voice": a voice that does not match any of the reciters it learned from.

Code, setup and full model card: https://github.com/aminkvh/Ganj_voice

Which files do I want?

You want to Use Where
read with the best quality, in Python with a GPU the full model (64 steps, several takes per line) this folder's top level and g2p/; GitHub release v1.2.0 (ganj-zip-d4h.zip)
read on a phone or laptop CPU, as the Ganj app does the distilled model (4 to 8 steps, ONNX) ganj-zip-distill-1.2.0/; GitHub release phone-1.2.0

Older app folders stay online unchanged, because installed apps download from them: ganj-zip-distill/ and the release phone-v1 (voice 1.1.0, used by Ganj 1.6.0), and the release v1.1.0 (ganj-zip-distill.zip, the same files as one zip). Version 1.2.0 brought a better letters-to-phonemes model (g2p/); the voice itself is unchanged since 1.1.0.

Files (full model)

File What it is
iter-40000-avg-2.pt, model.json the acoustic model (ZipVoice, about 123 M parameters, trained from scratch)
tokens.txt the model's phoneme and pause tokens
prompt.wav, prompt_phonemes.txt the voice prompt and its transcript
g2p/ the letters-to-phonemes model used for verse (a fine-tune of Homo-GE2PE; see g2p/NOTICE.md)
ezafe.joblib, vowels.json checks used to choose the best of several takes
example_in.jsonl, example.tsv, example.jsonl a worked example

Use

With the Ganj voice repository installed and a ZipVoice checkout set up (see its recipes/README.md), from the repository root:

python tools/reading_list.py ganj-zip-d4h/example_in.jsonl lists --voice ganj-zip-d4h
VOICE_DIR=ganj-zip-d4h recipes/zipvoice/synth_voice.sh lists/list.tsv out lists/text.jsonl

out/ then holds one 24 kHz WAV file per line. To read your own text, write it like example_in.jsonl: one JSON object per line with key, lines (the hemistichs, or one prose sentence), kind (verse or prose) and an optional metre. Hemistichs that are not Persian are left out, because the voice reads them badly.

On a phone: ganj-zip-distill-1.2.0/

The same voice, distilled so that it reads in 8 steps instead of 64, as ONNX files for sherpa-onnx 2.0.1 on a phone's CPU: the acoustic model at full precision (495 MB), the vocoder (54 MB), the voice prompt prepared for sherpa, and phone.json with the settings for sherpa and the clean-up after it. The folder's README gives the steps.

Training data

Ganjoor recitations (152.1 h, used with Ganjoor's consent) and ManaTTS prose (64.1 h, CC0). The voice prompt is built from a recitation by امیرحسین عظیمی, which is not in the training data. On 50 held-out lines the voice's highest speaker similarity to any training reciter is 0.49 (the limit for "nobody's voice" is 0.6).

Licence

Copyright (c) 2026 Amin Akbari. CC BY-NC 4.0 (LICENSE): free for non-commercial use with attribution. Commercial use needs a paid licence: COMMERCIAL.md. Credits: CREDITS.md.

Not for imitating real people: do not re-tune the voice to match a reciter.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support