Instructions to use venkycs/laya-telugu with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Laya
How to use venkycs/laya-telugu with Laya:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Laya (లయ)
Laya (లయ) is Telugu and Sanskrit for rhythm, tempo and the beat that holds speech and music together. The mark is that beat: a pulse that rises and settles under the word.
Laya is a small Telugu voice for casual, everyday Telugu with English words mixed in (Tinglish), the way people actually talk: "One minute, నేను ఇప్పుడే check చేస్తాను." It has 22.5 million parameters, runs on a laptop CPU, and has two voices, a female and a male. Private for now; an independent project by venkycs.
Status. Version 0.2, 16 Nov 2026. Clear on short everyday sentences; about as accurate as the voice it started from (1.8% vs 2.1% read-back error), not clearly better to a listener. It has no working emotion control. Read "What works and what does not" before you rely on it.
Quick start
pip install torch numpy safetensors huggingface_hub
huggingface-cli download venkycs/laya-telugu --local-dir laya-telugu && cd laya-telugu
python speak.py --text "Hi, నేను లయని. మీతో తెలుగులో chat చేయడం నాకు చాలా happy గా ఉంది." --out hello.wav
python speak.py --voice male --text "మీ reports అన్నీ normal గా ఉన్నాయి, tension పడకండి." --out doctor.wav
From Python (same folder):
from speak import load, say
tts, model = load()
say(tts, model, "Evening కలుద్దామా? ఒక coffee తాగుదాం.", "hello.wav", voice="female")
Options: --voice female|male, --style read|conv (conv is a little more conversational), --speed (1.15 is Laya's pace; 1.0 is calmer), --tail-ms (closing pause, default 300). speak.py first cleans the text with laya_tokenizer.py: English words, numbers, dates, casual punctuation and Telugu typed in English letters become Telugu script the voice can say (tension → టెన్షన్, 2:30 → రెండు ముప్పై). --raw skips that step. If a word is said wrongly, add its Telugu spelling to laya_lexicon.json.
Listen: one voice, many uses
Five longer samples in samples/usecases/, in casual Telugu with everyday English words. The scripts are in samples/usecases/scripts.json.
| Sample | Listen | Voice | Length | What it shows |
|---|---|---|---|---|
doctor.wav |
male, slow and calm (speed 1.0) | 40 s | Reassuring advice, medicine timing, numbers | |
call-center.wav |
female, brisk (1.1) | 43 s | Apology, order lookup, time and discount, English words | |
teacher.wav |
female, warm (1.05) | 49 s | Explaining a lesson, praise, questions | |
girlfriend.wav |
female, conversational style (1.1) | 36 s | Casual, affectionate chat | |
rasi-phalalu.wav |
male, steady (1.0) | 74 s | Weekly రాశి ఫలాలు for all twelve signs (for entertainment) |
The feeling in these samples comes from the wording, pace, pauses and choice of voice. Laya has no working emotion control, so it cannot sound truly happy, sad or angry on its own.
Short sentences
| Sentence | Listen |
|---|---|
| Hi, నేను లయని. మీతో తెలుగులో chat చేయడం నాకు చాలా happy గా ఉంది. | |
| మీ reports అన్నీ normal గా ఉన్నాయి, tension పడకండి. | |
| One minute, నేను ఇప్పుడే check చేస్తాను. | |
| Super, చాలా బాగా చెప్పావు! | |
| Evening కలుద్దామా? ఒక coffee తాగుదాం. | |
| మేష రాశి వాళ్ళకి, ఈ week work లో speed పెరుగుతుంది. | |
| నువ్వు ఎక్కడ ఉన్నావ్? నేను ఇంటికి వస్తున్నా. | |
| రేపు morning పది గంటలకి meeting ఉంది, మర్చిపోకు. |
The approach, at a glance
- Start from a small, fast architecture. A Matcha-TTS acoustic model (flow matching, 22.5M parameters) with a Vocos vocoder, pretrained on studio Telugu read speech by two speakers (Rasa). It reads one character at a time from a fixed set of 78 symbols.
- Fix the text, not the symbols. Casual Telugu is full of English words, numbers, dates, emoji and Telugu typed in English letters, which the model cannot say. We keep its symbol set (changing it would throw away what it learned) and clean the text in front of it, driven by an editable word list (
laya_lexicon.json) that a Telugu speaker can fix in one line. - Fine-tune in rounds on a laptop. Each round changes one thing and is scored on the same 100 held-out casual sentences: a teacher model transcribes our audio and we measure how much it differs from the input (read-back error), plus blind listening and a human rating sheet.
- Add casual Telugu. Round 2 added generated everyday sentences and real conversational Telugu speech, and a conversational style.
- Average back. Fine-tuning made the plain voice slightly worse, so the released Laya is a 50/50 average of the fine-tuned weights and the starting weights, with no further training. That gave the best accuracy we measured.
What we tried and dropped: emotion styles (three attempts; a judge recognised the intended emotion only 3 to 7% of the time) and a Telangana-accent version (Laya-TG; it did not sound natural, 1.75 out of 5, and was withdrawn).
Results
Read-back error on 100 casual Telugu-English sentences never used for training, female voice (lower is better; a gap under half a point is noise):
| Starting voice | Laya | |
|---|---|---|
| All 100 sentences | 2.1% | 1.8% |
| Sentences with dialect words | 5.6% | 3.7% |
| Sentences over 30% error | 0% | 0% |
Human ratings, one listener, 8 sentences, versions hidden, scores 1 to 5: Laya 3.38, starting voice 3.12.
Everything is in eval/ (results.json, human-ratings.json, readback-scores.json, and the emotion-judge result).
What works and what does not
- Works: short everyday Telugu with English words, numbers, times, dates and places; the closing pause; the text front end and its word list.
- Check the output: long, rare or technical English words can be misspelled by the front end and said wrongly; fix them in
laya_lexicon.json. Telugu typed fully in English letters ("nuvvu ekkada unnava") is read poorly. - Does not work: emotion styles (
happy,angry,sad,fear,disgust,surprise) and thetelanganastyle. They are still in the model's style table from training, butspeak.pyblocks them unless you pass--experimental.
Training data and credit
All data is in the private dataset venkycs/laya-telugu; see its SOURCES.md for sources, licences and consent.
- Rasa (AI4Bharat, CC BY 4.0): studio Telugu speech used for pretraining and fine-tuning. Please cite:
@inproceedings{ai4bharat2024rasa,
author={Praveen Srinivasa Varadhan and Ashwin Sankar and Giri Raju and Mitesh M. Khapra},
title={{Rasa: Building Expressive Speech Synthesis Systems for Indian Languages in Low-resource Settings}},
year=2024,
booktitle={Proc. INTERSPEECH 2024},
}
- Architecture and pretrained weights: Matcha-TTS (MIT) with the Vocos vocoder
charactr/vocos-mel-24khz(MIT). Laya's starting weights are a Telugu Matcha-TTS model by Sulabh Katiyar (CC BY 4.0, code MIT); this credit is required by its licence. - Vaani (IISc and ARTPARK, CC BY 4.0): Telugu clips, some with human transcripts and some transcribed by Gemini 3.1 Pro and Gemini 3.8 Flash.
- Podcast audio: author-attested ownership and consent. Used only for training; no podcast audio is redistributed here or in the dataset.
- Generated sentences: written by Gemini 3.1 Pro and spoken by the starting voices, marked as generated.
Limits and responsible use
- Two synthetic voices (one female, one male) from studio speakers. It does not clone anyone; do not use it to impersonate a real person or to mislead listeners about who is speaking.
- Short sentences work best. Check long, numeric or code-mixed text.
laya_tokenizer.unsayable()lists characters the voice cannot say. - The rasi phalalu sample is for entertainment and the doctor sample is a demo, not medical advice.
- Evaluated by a teacher model, one listener and 100 sentences. Do not read small differences as real.
Before making this public
- Confirm in writing the consent for the podcast audio used in training.
- Have more Telugu speakers rate the samples.
- Confirm the licence choice with the upstream author's terms (this card keeps CC BY 4.0 with full credit).
Files
| Path | What |
|---|---|
laya-ap/ |
The model (acoustic weights, vocoder, config, code, upstream licence) |
speak.py |
Command line and Python load() / say() |
laya_tokenizer.py, laya_lexicon.json |
The text front end and its word list |
samples/ |
usecases/ (five longer samples) and side-by-side/ (eight short sentences) |
logo/ |
laya-logo.svg (wordmark), laya-mark.svg (the beat), laya-icon.svg (app icon); the Telugu is outlines, so no font is needed |
eval/ |
Read-back scores, human ratings, emotion-judge result |
LICENSE.md |
Licence and credit |
Version history: 0.1 (16 Nov 2026, dataset only), 0.2 (16 Nov 2026, first model; a Telangana version was tried and withdrawn the same day).
- Downloads last month
- -