Downloads · 30 days
20
8% of all-time downloads
harinpurumandla/tenglish-to-telugu
tenglish-to-telugu is a machine learning model from harinpurumandla. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Converts colloquial Romanized Telugu (Tenglish) to Telugu Unicode script.
Downloads · 30 days
20
8% of all-time downloads
All-time downloads
248
Public
Parameters
300M
1.2 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors1.2 GB · 100%
From the Hugging Face model README
Converts colloquial Romanized Telugu (Tenglish) to Telugu Unicode script.
Fine-tuned from google/byt5-small on 5.6M transliteration pairs including 6,364 targeted synthetic pairs generated to fix colloquial error patterns identified in v1.0.
| Input (Tenglish) | Output (Telugu) | Notes |
|---|---|---|
| nenu telugu matladatanu | నేను తెలుగు మాట్లాడతాను | standard sentence |
| ela unnav bro | ఎలా ఉన్నావ్ బ్రో | casual greeting with English code-mix |
| chala bagundi aa movie | చాలా బాగుంది ఆ మూవీ | informal review |
| repu veltunnanu | రేపు వెళ్తున్నాను | near-future plan |
| antey em cheppali | అంటే ఏం చెప్పాలి | disambiguation: antey = "means" |
| ante idi chalu | అంతే ఇది చాలు | disambiguation: ante = "that's all" |
| naaku teliyadu | నాకు తెలియదు | dative suffix: naaku |
| pakka correct ga cheppadu | పక్కా కరెక్ట్ గా చెప్పాడు | Urdu loanword: pakka |
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("harinpurumandla/tenglish-to-telugu")
model = AutoModelForSeq2SeqLM.from_pretrained("harinpurumandla/tenglish-to-telugu")
def to_telugu(text):
inputs = tokenizer(text, return_tensors="pt", max_length=128, truncation=True)
outputs = model.generate(**inputs, max_length=128, num_beams=4)
return tokenizer.decode(outputs[0], skip_special_tokens=True)
print(to_telugu("nenu telugu matladatanu"))
# నేను తెలుగు మాట్లాడతాను
print(to_telugu("ela unnav bro"))
# ఎలా ఉన్నావ్ బ్రో
| Property | Value |
|---|---|
| Base model | google/byt5-small |
| Architecture | Byte-level T5 (seq2seq) |
| Parameters | ~300M |
| Direction | Tenglish (Roman) → Telugu script |
| Training data | 5,586,373 pairs (5 corpus sources + 6,364 synthetic) |
| Training | 1 epoch continued fine-tune, LR=1e-5, bf16, dual RTX 3090 |
| Eval CER | 1.89% |
| Decoding | Beam search (num_beams=4) |
| Max input length | 128 tokens (bytes) |
| License | Apache-2.0 |
Evaluated on a held-out val set (5,000 sampled pairs), beam search decoding (num_beams=4).
| Version | CER | Notes |
|---|---|---|
| v1.0 | 7.08% | Baseline — trained from scratch on 5.5M pairs |
| v1.1 | 1.89% | Continued fine-tune with targeted synthetic colloquial data |
Evaluated on 50 clean Tenglish inputs (8–60 chars, alphabetic):
| Model | CER | Parameters |
|---|---|---|
| ByT5-small v1.1 (this model) | 2.70% | 300M |
| Gemma4 31B (zero-shot, Ollama) | 8.48% | 31B |
ByT5-small achieves 3× lower error rate at 1% of the parameter count. The key difference: Gemma4 normalizes colloquial spellings toward standard Telugu (e.g., chupiyaru → చూపించారు), while this model faithfully preserves the informal romanization convention (→ చుపియారు). For Tenglish transliteration, faithfulness to the input spelling is the correct behavior.
| Slice | v1.0 CER | Coverage |
|---|---|---|
| Overall | 7.08% | 10,000 pairs |
| Formal | 3.25% | 71 pairs |
| Colloquial | 3.09% | 3,656 pairs |
| Long sentence | 16.21% | 206 pairs |
v1.1 added 6,364 targeted synthetic pairs generated with local Ollama gemma4:31b covering 9 confusion categories identified from v1.0 error analysis:
| Category | Pairs | Focus |
|---|---|---|
| antey_disambiguation | 704 | అంటే (means) vs అంతే (that's all) in context |
| dative_suffix_context | 701 | naaku / naku / ki suffix forms |
| colloquial_chat | 705 | WhatsApp-style greetings, questions, reactions |
| urdu_loanwords | 1,000 | pakka, bilkul, mast, zabardast in Telugu sentences |
| code_mix | 1,000 | English words with Telugu morphological markers |
| long_sentences | 800 | Full sentences >15 words |
| reactions_expressions | 701 | Emotional interjections and expressive phrases |
| continuous_negation | 403 | Negation + continuous tense combinations |
| long_vowel_disambiguation | 350 | Phonemically distinct long vs short vowels |
All synthetic pairs were validated for Telugu script density (≥10%), length ratio, and ASCII content before inclusion.
For the reverse direction (Telugu script → Tenglish), see harinpurumandla/telugu-to-tenglish.
@misc{telugu-transliterator-2026,
title={Telugu Transliterator: Tenglish to Telugu Unicode Transliteration},
author={Purumandla, Sai Harin Kumar Reddy},
year={2026},
url={https://huggingface.co/harinpurumandla/tenglish-to-telugu}
}