Urdu Male Voice — Piper TTS (medium)

Urdu text-to-speech voice for Piper.

Overview

• Language: Urdu (ur_PK)
• Voice: Male
• Quality: Medium
• Phonemizer: eSpeak (ur)
• Sample rate: 22050 Hz
• Current version: v2 (October 2026)

Listen

Optimized Inference Parameters

Usage

echo 'یہ ایک سادہ جملہ ہے۔' | piper \
  --model ur_PK-fasih-medium-model.onnx \
  --config ur_PK-fasih-medium-model.onnx.json \
  --output_file output.wav

Training Details

v2 (October 2026):

• Base model: v1 of this voice (epoch=3206-step=383452.ckpt), full-model fine-tune
• Dataset: Urdu recordings by Muhammad Noman (same speaker), 11,998 curated clips (~15 h) drawn from his speech collection (Urdu_Audio-Realistic), used with the contributor's approval
• Training: steps 383,452 -> 473,452 on an Apple M4 Mac mini (CPU, batch 8)
• Final checkpoint: checkpoint/epoch=3240-step=473452.ckpt (val_mel 0.4129)
• Listening check: 3.3/5 vs 2.1/5 for v1 on the same 10 Urdu sentences (blind re-scoring by the CSUN 2027 paper's Urdu reviewer)

v1:

• Base model: Hindi (hi_IN) medium (Rohan) checkpoint
• Base checkpoint: epoch=3190-step=309852.ckpt https://huggingface.co/datasets/rhasspy/piper-checkpoints/tree/main/hi/hi_IN/rohan/medium
• Dataset: UAT (Urdu)
• Training checkpoint: epoch=3206-step=383452.ckpt
• Hardware: Apple M4 Mac mini

Notes

• Trained on Urdu dataset with eSpeak phonemization.
• Based on Hindi model initialization, adapted for Urdu speech.
• Quality may vary on mixed-language input (Urdu + English).

Links

• Piper — Neural TTS engine

License

MIT