Lafzyn

Lafzyn is a fine-tuned Qwen3.5-0.8B model that converts Urdu script into IPA phonetic transcription.

Model Details

Property	Value
Base model	Qwen3.5-0.8B
Task	Urdu Grapheme-to-Phoneme (G2P)
Output format	IPA transcription
Training data	100,000 Urdu IPA pairs

---

Evaluation Results

Evaluated on 500 held out samples from the phoneme map using Phoneme Error Rate (PER) character level edit distance on IPA strings (lower is better).

Metric	Value
Mean PER	16.94%
Perfect transcriptions (PER = 0)	88 / 500 - 17.6%
Near perfect (PER < 10%)	166 / 500 - 33.2%

Sample Outputs

Best predictions (PER = 0.000):

Urdu	Expected IPA	Predicted IPA
مختار احمد	mʊxˈt̪aːr ˈæhməd	mʊxˈt̪aːr ˈæhməd
قابل رحم	qɑːˈbɪl rɛˈhəm	qɑːˈbɪl rɛˈhəm
رواں ہفتے	rəˈʋãː ˈhəft̪eː	rəˈʋãː ˈhəft̪eː
دفاعی کمیشن	d̪ɪˈfaː.iː kəˈmɪ.ʃən	d̪ɪˈfaː.iː kəˈmɪ.ʃən
نفیس المزاج	nəˈfiːs əlmɪˈzaːd͡ʒ	nəˈfiːs əlmɪˈzaːd͡ʒ

Hardest cases (Arabic loanwords & rare compounds):

Urdu	Expected IPA	Predicted IPA	PER
ائمہ سبعہ	aɪˈʔɪm.maː ˈsab.ʕa	ˈɛːmɑː sɪˈbɑː	0.722
نفع المصرف	ˈnafʕaː alˈmasˤrif	nəˈfɑːl ʔalˈmɪrɑːf	0.667
میٹر کیولیٹ	ˈmeːʈər ˈkjuːlɪt	meːʈər kiːˈloːɛt̪	0.562

High error cases are predominantly rare Arabic origin compound words and technical transliterations that are underrepresented in training data.

---

Quick Start

from transformers import AutoTokenizer, pipeline
import torch

pipe = pipeline(
    "text-generation",
    model="mahwizzzz/qwen3.5-g2p",
    torch_dtype=torch.float16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are an expert Urdu linguist and phonetician. Convert the given Urdu word or phrase into its IPA transcription. Output only the IPA string, nothing else."},
    {"role": "user",   "content": "پاکستان"},
]
out = pipe(messages, max_new_tokens=64, do_sample=False)
print(out[0]["generated_text"][-1]["content"])  # → pɑːkɪsˈt̪aːn

---

Phoneme Inventory

The model covers the full standard Urdu IPA phoneme set:

• Stops: b p t̪ ʈ d̪ ɖ k ɡ q ʔ
• Fricatives: f s ʃ z ʒ x ɣ ɦ h ħ ʕ
• Affricates: tʃ dʒ
• Nasals: m n ɳ ŋ n̪
• Liquids & glides: r ɾ ɽ l w j ʋ
• Vowels: ə ɪ ʊ aː iː uː eː oː ɛ ɔ æ
• Diacritics: ː (length) ̃ (nasalization) ˈ ˌ (stress)

---

Limitations

• Arabic origin loanwords and rare compounds show higher error rates (PER > 0.5)
• Optimised for Modern Standard Urdu; regional dialects and heavy code-switching may degrade accuracy

---

Related Resources

Resource	Link
🤗 Live Demo	spaces/mahwizzzz/lafzyn
📦 GGUF (llama.cpp / Ollama)	mahwizzzz/lafzyn-gguf
🏋 Base model	Qwen/Qwen3.5-0.8B

---

Citation

@misc{lafzyn2026,
  title   = {Lafzyn: Urdu Grapheme-to-Phoneme with Qwen3.5-0.8B},
  author  = {Mahwiz Khalil},
  year    = {2026},
  url     = {https://huggingface.co/mahwizzzz/lafzyn},
  note    = {Fine-tuned on 100k Urdu IPA pairs, mean PER 16.94\%}
}