nemo-sindhi-g2p-conformer

Sindhi Conformer G2P

Sindhi Conformer G2P converts Pakistani Sindhi-script text into broad IPA-style phoneme strings. It is a compact character-level Conformer CTC model built for TTS preprocessing, pronunciation-label bootstrapping, and speech-dataset quality checks.

This model belongs to the Pakistani Languages G2P Family alongside the Pashto Conformer G2P. The family is designed to grow into a consistent set of practical G2P models for Pakistan's languages.

Model Files

Quick Start

Install the dependencies:

pip install -r requirements.txt

Run from a cloned repository:

python inference.py --text "سنڌي ٻولي اسان جي سڃاڻپ آهي"

The same script can download the model directly from Hugging Face when the local .nemo file is absent:

python inference.py --repo-id TBOGamer22/nemo-sindhi-g2p-conformer --text "تعليم انسان جي زندگي بدلائي ٿي"

Batch a UTF-8 text file containing one sentence per line:

python inference.py --text-file inputs.txt --batch-size 32 --json

Force CPU inference:

python inference.py --device cpu --text "هي هڪ آزمائشي جملو آهي"

Architecture

• Framework: NVIDIA NeMo
• Model family: Conformer CTC G2P
• Input: Pakistani Sindhi-script characters
• Output: broad IPA-style phoneme characters
• Encoder layers: 8
• Model dimension: 192
• Attention heads: 4
• Convolution kernel size: 15
• Source repetition for CTC: 3
• Maximum repeated source length: 768

Evaluation

Training was intentionally stopped after the validation pass for epoch 11. The release uses the best fully saved checkpoint rather than the interrupted partial epoch. That checkpoint was evaluated on the held-out test set.

Selected checkpoint: sindhi-conformer-g2p-epoch=11-val_per=0.0113.ckpt

PER is character-level Levenshtein distance divided by reference IPA-character count. Weighted PER aggregates edits and reference units across the complete test split.

Intended Use

• Pakistani Sindhi text-to-phoneme generation
• TTS text preprocessing
• Pronunciation label bootstrapping
• G2P consistency checks and speech-corpus QA

This model is not an ASR system, general language model, translator, dialect classifier, or narrow phonetic transcriber.

Limitations

• Outputs are broad, corpus-style IPA rather than narrow phonetic transcriptions.
• Names, abbreviations, numbers, punctuation-heavy text, rare spellings, and code-switching can be difficult.
• Pakistani Sindhi pronunciation and spelling variation is not fully represented.
• The model predicts from text only and does not listen to audio during inference.
• Very low in-domain PER may not transfer unchanged to unrelated corpora or dialects.

Ethical and Practical Notes

Review outputs before using them for linguistic claims or high-stakes applications. TTS and corpus-building workflows should retain human review for names, dialectal vocabulary, and code-switched utterances.

License

This model is released under the MIT License.

Author

Developed and released by:

Talha Bin Omar

Citation

If you use this model, please cite it as:

@misc{binomar2026sindhi_conformer_g2p,
  title        = {Sindhi Conformer G2P: An Open-Source Grapheme-to-Phoneme Model for Pakistani Sindhi},
  author       = {Bin Omar, Talha},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/TBOGamer22/nemo-sindhi-g2p-conformer}},
  note         = {Conformer CTC G2P model for broad IPA phoneme generation}
}