An open voice for Tamil.
Geetha turns Tamil text into speech in two studio voices. She can sound happy or afraid, and she can take a breath before she starts. Everything she is made of is public.
வணக்கம், என் பேரு கீதா. நான் தமிழ்ல பேசுற ஒரு குரல்.
Hello, my name is Geetha. I'm a voice that speaks Tamil.
Anitha, one of two voices
Two voices, Anitha and Arun. Six moods to speak in, and a breath before she begins.
Everyday Tamil, the way it is spoken
These are selected takes, picked from up to three tries each. Pick a voice and press play.
Tell her how to feel
One word changes the delivery. It works best when the sentence already carries the feeling. Anitha speaks all of these.
You write a line. She reads it.
One Python file is the whole recipe. Name the voice, add a style if you want one, and write [breath] where she should breathe in first. Numbers, dates and rupees are read as Tamil words.
from inference import Geetha
tts = Geetha("Dhevenddra/geetha-tts")
The breath is placed in, not learned. The studio recordings behind these two voices hold no breaths, and teaching it twice did not take. So [breath] writes one of her own breaths into the stream and she carries on from it.
Measured, and said as measured
304 Tamil sentences, one take each, heard by two automatic judges. The number is the share of characters the judge got wrong, so lower is better.
Character error, Indic-Transcribe-core judge. The second judge, IndicConformer, gives 2.85, 2.60, 8.76 and 8.33.
Level with her base, not better
The fine-tune did not make the words clearer. The gap to Indic-Speak is a quarter of a point and the interval crosses zero.
She stops when the sentence does
The base ran on to the 22-second cap on 20 of 304 sentences. Geetha did on 1.
The moods are audible
In 24 blind pairs the styled take was picked every time. That is one listener who knows the voices, not a panel.
What she can't do yet
- She is slow: about 40 seconds for a 7-second sentence on a free-tier GPU.
- English words inside Tamil sentences are her weak spot.
- She sounds like a clear reader in a studio, not a person talking in a room.
- No voice cloning, no laughter or sighs, no other languages.
Short answers
What is Geetha?
Geetha is an open text-to-speech model for Tamil. It is a LoRA fine-tune of Indic-Speak by Bodhan AI, trained on about 126 hours of openly licensed Tamil speech. It speaks in two studio voices, Anitha and Arun.
How accurate is Geetha?
On 304 Tamil sentences Geetha scores 2.15% and 2.85% character error with two automatic judges. That is level with its base model, not better. Two other open Tamil models scored between 5% and 9% on the same sentences.
Can Geetha express emotion?
Geetha has six styles: happy, sad, angry, afraid, surprised and disgusted. In 24 blind pairs one listener picked the styled take every time. Emotion is clearest when the sentence itself carries the feeling.
Does Geetha really breathe?
The breath is placed in, not learned. Writing [breath] at the start of a sentence puts one of the model's own breaths into the audio and the voice carries on from it. A breath in the middle of a sentence is experimental and off by default.
How do I run Geetha?
Ask for access to the gated base model on Hugging Face, log in, download three small files from the Geetha model page and call tts.speak with your Tamil text. A free-tier T4 GPU is enough. Expect 15 to 40 seconds for a sentence.
Is Geetha free to use?
Yes. The code is Apache-2.0 and the weights follow the base model's Indic Open Model License v1.0, which allows commercial use with attribution. You need access to the gated base model to run it.
Who made Geetha?
Geetha was built by Dhevenddra in Tamil Nadu, on Indic-Speak by Bodhan AI and openly licensed speech from AI4Bharat. The model, the code and all 912 benchmark clips are public.
None of this asks for your trust
All 304 sentences, what every system said, what both judges heard, the clips that went wrong and the code that scored them are public.
Built on Indic-Speak by Bodhan AI, with openly licensed speech from IndicVoices-R and Rasa and judge models from AI4Bharat. Thank you to the teams behind Rumik-OSS-1 and snorTTS-Indic for releasing open Tamil models to measure against.