kurmanci-tts-piper

Kurmancî TTS — Piper ONNX

Kurmancî TTS is a multi-speaker text-to-speech model for Kurmanji (Northern Kurdish). It was fine-tuned with the Piper training pipeline and exported to ONNX for fast, CPU-friendly inference.

Try it in the interactive Kurmancî TTS Space.

Model summary

The repository contains the two files required by Piper:

• model.onnx — exported acoustic model.
• model.onnx.json — phoneme, speaker, audio, and inference configuration.

Usage

Install Piper and download the model files:

pip install piper-tts huggingface_hub
hf download atenax/kurmanci-tts-piper \
  model.onnx model.onnx.json --local-dir ./kurmanci-tts

Generate a WAV file with the benchmark-selected speaker ID 54:

echo "Roj baş! Tu çawa yî?" | piper \
  --model ./kurmanci-tts/model.onnx \
  --config ./kurmanci-tts/model.onnx.json \
  --speaker 54 \
  --output_file kurmanci.wav

Use an integer from 0 through 89 to explore the learned speaker identities. Speaker 54 is the recommended general-purpose starting point; it was selected after comparing every speaker on short and extended Kurmanji test phrases. For the simplest browser experience, use the linked Space.

Optional clarity cleanup

The Space enables a conservative browser-local clarity filter by default. It removes sub-bass rumble, gently reduces high-frequency sizzle, and preserves a raw-output option. The same cleanup can be applied to CLI output with FFmpeg:

ffmpeg -i kurmanci.wav \
  -af "highpass=f=65,equalizer=f=6500:t=q:w=1:g=-4,lowpass=f=9500" \
  -c:a pcm_s16le kurmanci-clean.wav

This filter is intentionally light. Strong noise suppression can introduce metallic artifacts and cannot correct pronunciation or vocoder limitations.

Training data

The model was prepared from the atenax/kmr_tts Kurmanji speech corpus. The processed training metadata contains 56,499 utterances across 90 speaker identities. See the dataset card for provenance and preparation notes.

Intended use

This model is intended for:

• Kurmanji language technology research and education.
• Accessibility and reading-assistance prototypes.
• Local or server-side speech synthesis experiments.
• Applications where generated audio is reviewed before publication.

Limitations and responsible use

• Pronunciation, prosody, and speaker consistency can vary by sentence and ID.
• Quality may decrease for very long text, code-switching, unusual names, or writing outside the Kurmanji Latin alphabet.
• Speaker IDs are anonymized learned identities, not named or verified people.
• Do not use the model for impersonation, fraud, harassment, or misleading synthetic media. Disclose generated speech when context requires it.

Reproducibility

The configuration records the recommended default (default_speaker_id=54), the inference defaults (noise_scale=0.52, length_scale=1.02, noise_w=0.58), and the full phoneme and speaker mappings. Both files should be kept together when redistributing or deploying the model.

Acknowledgements

Built with the open-source Piper speech synthesis project and eSpeak NG.