🇮🇷 Persian TTS — Piper EN Base → ManaTTS (v1)

Model name: fa-ir-tts-piper-en-mantatts-v1
Previous name: kiarashQ/fa_IR-mantatts
Sampling rate: 22,050 Hz
Base checkpoint:
ar/ar_JO/kareem/medium/epoch=5079-step=1682020.ckpt (Piper AR, medium)

This is a Persian (fa-IR) single-speaker TTS model fine-tuned from the Arabic Piper medium checkpoint on the ManaTTS dataset.

---

⭐ Highlights

• ✔️ Arabic phoneme system provides better accuracy for certain Persian words
• ✔️ Produces stable, smooth speech
• ✔️ Complements the EN-based model — each excels at different phonemes
• ✔️ Output at 22.05 kHz

---

🧪 Training Details

Training script: piper_train
Hardware: 1× GPU A4000
Dataset: ManaTTS
Batch size: 16
Precision: 32-bit
Validation split: 1%
Test samples: 5
Training epochs: 20
Logging: every 2000 steps
Quality setting: medium
Checkpoint frequency: every 1 epoch
No resume checkpoint (fresh fine-tune)

Training command:

piper_train \
  --dataset-dir /workspace/piper_full/piper_dataset \
  --accelerator gpu --devices 1 \
  --batch-size 16 \
  --validation-split 0.01 \
  --num-test-examples 5 \
  --quality medium \
  --checkpoint-epochs 1 \
  --max_epochs 20 \
  --precision 32 \
  --log_every_n_steps 2000

🔊 Inference Example

piper \
  --model model.onnx \
  --config config.json \
  --text "سلام! حال شما چطور است؟" \
  --output_file out.wav

Python:

import subprocess

text = "سلام! امروز هوا چطور است؟"
subprocess.run([
    "piper", "--model", "model.onnx", "--config", "config.json",
    "--text", text, "--output_file", "out.wav"
])

🔍 Observations

• The Arabic-base version sometimes pronounces Persian words more correctly than the EN-base model.
• Slightly lower overall accent naturalness compared to EN-base.
• Useful as a complementary voice to EN-base.

📜 License

Apache-2.0

🙏 Credits

• Piper TTS
• ManaTTS dataset
• Model fine-tuning by @kiarashQ