ur_PK-aegis_female-medium
Female Urdu speech-synthesis voice in Piper ONNX format.
Purpose
CPU inference of a female ur_PK voice. The previous card describes the quality tier as medium and the sample rate as 22050 Hz, with an eSpeak-ng ur phonemizer.
| Language | Urdu (ur_PK) |
| Gender | Female |
| Quality | Medium |
| Sample rate | 22050 Hz |
| Phonemizer | eSpeak (ur) |
| ONNX file | ur_PK-aegis_female-medium.onnx |
| ONNX size | 63,515,589 bytes (Hub Content-Length on the default branch, retrieved 2 October 2026) |
The article title rounds the release to about 61 MB. The byte count above is the file size, not an audio-quality metric.
Upstream attribution
This card states that the voice was fine-tuned from Fasih (ur_PK-male-medium) by IhorShevchuk: https://huggingface.co/IhorShevchuk/piper-voice-ur-fasih
Fasih was itself trained from a Hindi (hi_IN) medium checkpoint (Rohan), via rhasspy/piper-checkpoints.
A separate article, Distilling OmniVoice into Aegis, describes data-level distillation from OmniVoice into a VITS student and describes Fasih as a complementary male voice. This repository does not contain the distillation corpus, teacher outputs, or training logs, and the two accounts are not reconciled here.
Data provenance
Not in this repository: training transcript source, hours, speaker count, whether the audio is recorded or synthesized, and the data license. The article linked above discusses distillation. Those details are not copied here as repository evidence.
Training
The previous card states:
- Fine-tuned from the Fasih Urdu male checkpoint.
- Phonemes: eSpeak-ng
urIPA.
Steps, learning rate, hardware, and seed are not in this repository.
Evaluation
No word-error rate, MOS, or speaker-similarity figure is stored in this repository. Audio samples are under samples/.
Limitations
- Quality, dialect coverage, and code-switching behavior are not measured in this card.
- The fine-tune-from-Fasih statement and the distillation article are unreconciled.
- eSpeak Urdu phonemization can diverge from native pronunciation. That is a property of the named phonemizer, not a new measured error rate.
License
MIT. Copyright (c) 2026 Muhammad Mahwiz Khalil (Proxima AI).
This model is licensed under MIT, which requires retaining this copyright and license notice in all copies or substantial portions of the work, including redistributions and re-uploads.
If you use, fine-tune, or redistribute this model, retain attribution to:
Muhammad Mahwiz Khalil (Proxima AI) — https://huggingface.co/mahwizzzz
See the LICENSE file. The copyright notice must be preserved in any copy or redistribution.
Usage
echo 'یہ ایک سادہ جملہ ہے۔' | piper \
--model ur_PK-aegis_female-medium.onnx \
--config ur_PK-aegis_female-medium.onnx.json \
--output_file output.wav
Files:
ur_PK-aegis_female-medium.onnx
ur_PK-aegis_female-medium.onnx.json
samples/speaker_0.mp3
Citation
@misc{aegis_urdu_tts_2026,
author = {Khalil, Muhammad Mahwiz},
title = {Aegis: Female Urdu Piper TTS Voice (ur_PK-aegis_female-medium)},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/mahwizzzz/piper-voice-ur-aegis-female}
}