Downloads · 30 days
9
6% of all-time downloads
harinpurumandla/telugu-to-tenglish
telugu-to-tenglish 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 Telugu Unicode script to colloquial Romanized Telugu (Tenglish) — the reverse direction of harinpurumandla/tenglish-to-telugu.
Downloads · 30 days
9
6% of all-time downloads
All-time downloads
140
Public
Parameters
300M
2.4 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors1.2 GB · 100%
From the Hugging Face model README
Converts Telugu Unicode script to colloquial Romanized Telugu (Tenglish) — the reverse direction of harinpurumandla/tenglish-to-telugu.
Fine-tuned from google/byt5-small on 5.6M transliteration pairs including 6,364 targeted synthetic pairs covering colloquial confusion patterns.
| Input (Telugu) | Output (Tenglish) | Notes |
|---|---|---|
| నేను తెలుగు మాట్లాడతాను | nenu telugu matladatanu | standard sentence |
| ఏమి చేస్తున్నావు | emi chestunavu | casual question |
| నువ్వు ఎలా ఉన్నావు | nuvvu ela unnavu | greeting |
| చాలా బాగుంది | chala bagundi | common phrase |
| రేపు కలుద్దాం | repu kaluddam | informal plan |
| సరే అలాగే చేస్తాను | sare alage chestanu | agreement |
| పక్కా చెప్తున్నా | pakka cheptunna | Urdu loanword |
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("harinpurumandla/telugu-to-tenglish")
model = AutoModelForSeq2SeqLM.from_pretrained("harinpurumandla/telugu-to-tenglish")
def to_tenglish(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_tenglish("నేను తెలుగు మాట్లాడతాను"))
# nenu telugu matladatanu
print(to_tenglish("చాలా బాగుంది"))
# chala bagundi
| Property | Value |
|---|---|
| Base model | google/byt5-small |
| Architecture | Byte-level T5 (seq2seq) |
| Parameters | ~300M |
| Direction | Telugu script → Tenglish (Roman) |
| 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 | 16.69% |
| 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 | 16.53% | Baseline — trained from scratch on 5.5M pairs |
| v1.1 | 16.69% | Continued fine-tune with targeted synthetic colloquial data |
The CER is essentially flat between versions — the synthetic colloquial data did not significantly change reverse model performance. This is expected: the reverse direction (Telugu → Tenglish) has inherent ambiguity that cannot be resolved by more training data alone. The same Telugu word has multiple valid Tenglish spellings (e.g., అంతే → anthe, anthey, antey, ante), so CER against a single reference underestimates true model quality.
The forward model (Tenglish → Telugu) achieves 1.89% CER because Telugu Unicode is phonetically consistent — each romanization maps to one correct script form. The reverse direction has no such constraint: Tenglish spelling is not standardized, so the model must guess which of several valid conventions the reference used. A prediction of nenu vs nennu is linguistically correct but scores as an error.
aa→a, uu→u).For the forward direction (Tenglish → Telugu script), see harinpurumandla/tenglish-to-telugu.
@misc{telugu-to-tenglish-2026,
title={Telugu to Tenglish: Reverse Transliteration for Telugu Unicode Script},
author={Purumandla, Sai Harin Kumar Reddy},
year={2026},
url={https://huggingface.co/harinpurumandla/telugu-to-tenglish}
}