Negara-G2P-v2 (Homograph-Disambiguated & Natural Pronunciation)
This model is an advanced, robust fine-tuned version of negara-g2p-clean-v2 specifically optimized for Persian Grapheme-to-Phoneme (G2P) conversion. It achieves state-of-the-art results in Contextual Homograph Disambiguation (رفع ابهام هم‌نگاشت‌ها) while consistently maintaining superior baseline phonetic precision without catastrophic forgetting.
📊 Scientific Benchmark & Evaluation
To evaluate the model rigorously without synthetic bias, we benchmarked the fine-tuned checkpoint against the base model using Cleaned Human Ground Truth (derived from the human-annotated HomoRich benchmark) alongside a specialized Homograph Disambiguation Test Suite (19 Contextual Pairs):
| Model Name | Homograph Accuracy (%) ⬆️ | True PER (%) ⬇️ | True WER (%) ⬇️ | Latency (ms) ⚡ |
|---|---|---|---|---|
Base Model (negara-g2p-clean-v2) | 63.16% | 2.59% | 8.62% | 145.40 ms |
| Our Fine-Tuned Model (Optimal) | 94.74% (+31.58%) | 2.41% (Best) | 8.33% (Best) | 175.89 ms |
🌟 Key Highlights:
- +31.58% Absolute Leap in Disambiguation: Resolves highly ambiguous Persian homographs depending on syntax and surrounding semantic context (e.g., maskan [home] vs. mosakken [painkiller], keSti [ship] vs. koSti [wrestling], deh [village] vs. dah [ten], mohr [stamp] vs. mehr [affection]).
- Zero Catastrophic Forgetting: Not only preserves base phonetic capabilities, but lowers the overall Phoneme Error Rate (PER: 2.41%) and Word Error Rate (WER: 8.33%) across general Persian text.
- Optimized for Persian TTS Pipelines: Generates clean, consistent phonetic notation ideal for acoustic models (such as VITS, FastSpeech2, and Tacotron2).
🔍 Qualitative Comparison (Base vs. Fine-Tuned)
| # | Input Sentence (Persian) | Base Model (negara-g2p-clean-v2) | Fine-Tuned Model (Ours) | Target Context / Meaning |
|---|---|---|---|---|
| 1 | کشتی بزرگی در دریا غرق شد. | kaStiye bozorgi... ❌ | keStiye bozorgi... ✅ | Ship (keSti) vs. Wrestling (koSti) |
| 2 | از مسکن بیرون آمدم تا یک قرص مسکن خوردم. | az mosakken ... qorse mosakken ❌ | az maskan ... qorse mosakken ✅ | Home (maskan) vs. Painkiller (mosakken) |
| 3 | فضای ده بسیار آرام بود و ده نفر بیدار بودند. | fazAye dah ... faqat dah nafar ❌ | fazAye deh ... faqat dah nafar ✅ | Village (deh) vs. Ten (dah) |
| 4 | به سمت مزرعه رفت، مردی که سمت کدخدایی داشت. | be samte ... samte kadkhodAi ❌ | be samte ... semate kadxodAyi ✅ | Direction (samt) vs. Job Title (semat) |
| 5 | حیوان درنده بود؛ شاید یک ببر! وسایلت را ببر! | ... SAyad yek bebar! ... bebar ❌ | ... SAyad yek babr! ... bebar ✅ | Tiger (babr) vs. To Take (bebar) |
| 6 | در کشور قطر بود و لوله‌ای به قطر چند متر ساخت. | keSvare qatar ... be qatar Cand metr ❌ | keSvare qatar ... be qotre Cand metr ✅ | Country Qatar (qatar) vs. Diameter (qotr) |
| 7 | او با کمال مهر به نامه‌ای مهر تایید زد. | kamAle mehr ... mehre tayid ❌ | kamAle mehr ... mohre ta'yid ✅ | Affection (mehr) vs. Stamp/Seal (mohr) |
| 8 | به شدت مراقب عرض و آبروی دیگران بود. | morAqebe arz va Abru ❌ | morAqebe erz o Abru ✅ | Honor/Dignity (erz) vs. Expressing (arz) |
💻 Usage Example
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
device = "cuda" if torch.cuda.is_available() else "cpu"
model_id = "jafarahmadi/negara-g2p-v2-homorich-disambiguated"
# Load tokenizer and fine-tuned model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id).to(device)
model.eval()
# Sample text with homograph challenges
text = "کشتی بزرگی در دریا غرق شد و مسافران از مسکن خود دور شدند."
inputs = tokenizer(text, return_tensors="pt").to(device)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_length=256,
num_beams=4,
early_stopping=True
)
phonemes = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(phonemes)
# Expected Output: keStiye bozorgi dar daryA qarq Sod va mosAferAn az maskane xod dur Sodand