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

Usage

python

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

Model Details

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

Evaluation

Evaluated on a held-out val set (5,000 sampled pairs), beam search decoding (num_beams=4).

Version History

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.

Why Reverse CER is Higher than Forward

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.

Known Limitations

Spelling convention: Multiple valid Tenglish spellings exist for most Telugu words — the model outputs one learned convention which may differ from a user's preferred spelling.
Vowel length: Long vowels (ఆ, ఈ, ఊ) may be rendered as short in output (aa→a, uu→u).
Suffix truncation: Final short syllables occasionally dropped on longer words.
Long sentences (>40 words): ByT5 1024-byte window truncates inputs. Split long text before transliterating.

Companion Model

For the forward direction (Tenglish → Telugu script), see harinpurumandla/tenglish-to-telugu.

Citation

bibtex

@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}
}