Shqip · English
Modeli i zërit i Llafazanit: Kokoro-82M i përshtatur për shqipen, me dy zëra, Agon dhe Dardan. Është i njëjti model që përdorin programi për Windows, Llafazani në shfletues dhe shtesa për shfletuesin.
albanian_text.py (numrat, datat, shkurtimet dhe akronimet shkruhen me fjalë, pastaj kthehen në fonema), siç bën lexo.py.lexo.py e lexon tekstin njësoj si Llafazani: me të njëjtin ritëm, të njëjtat pauza dhe të njëjtin tingull.pip install kokoro soundfile
python lexo.py teksti.txt dalja.wav agon # ose dardan; shpejtësia: 0.8 më ngadalë, 1.3 më shpejt
The voice model of Llafazani: Kokoro-82M fine-tuned for Albanian, with two voices, Agon and Dardan. It is the model used by the Llafazani Windows app, by Llafazani in the browser and by the Llafazani browser extension.
| File | What it is |
|---|---|
kokoro-shqip.pth |
the weights (PyTorch, the standard Kokoro format, 82 million parameters) |
config.json |
the model's configuration and its phoneme vocabulary |
voices/agon.pt, voices/dardan.pt |
the two voices |
albanian_text.py, albanian_g2p.py |
Albanian text to phonemes: numbers, dates, abbreviations and acronyms written out in words, sentences, then phonemes in the model's symbol set |
lexo.py |
reads a text into a WAV file the way the Llafazani apps read it |
onnx/ |
the same model in ONNX, as the Llafazani apps use it (a text graph and a decoder graph, the voices as .npy) |
pip install kokoro soundfile
python lexo.py text.txt output.wav agon # or dardan; optional speed: 0.8 slower, 1.3 faster
import soundfile as sf
from lexo import read
sf.write("output.wav", read("Mirë se vini në Llafazani!", "agon"), 24000)
albanian_text (as lexo.py does). One sentence at a time, up to about 480 phonemes.lexo.py takes the same steps as the Llafazani apps: every sentence is read after a short ghost word that is cut away afterwards (no hesitation at the start), one voice style for the whole text, pauses at ; and : made by the model itself, sentence levels evened out.lexo.py reads at Llafazani's natural pace, about 170 words a minute with the pauses between sentences. In plain Kokoro terms, that is a speed of 0.90 for Agon and 0.91 for Dardan.Agon (the default) and Dardan are synthetic male voices created for Llafazani; they are not the voice of any real person. Please do not present them as the voice of a real person.
Apache License 2.0, like Kokoro-82M by hexgrad, on which this model is based. Free to use, including for commercial purposes.
Contact: kontakt@llafazani.com · llafazani.com
kokoro-shqip.pth 327 MBonnx/llafazani_decoder.onnx 214 MBonnx/llafazani_text.onnx 112 MBvoices/agon.pt 523.9 kBvoices/dardan.pt 523.9 kBhexgrad/Kokoro-82M