Indic Tiny TTS: Kannada
A tiny Kannada text-to-speech model: 22.5 million parameters, a female and a male voice, reads Kannada script. On FLORES-200 news sentences it makes fewer recognition errors (1.1 % CER) than 5 other open Kannada models tested. It runs 38× faster than real time on one Apple M1 CPU thread and 7× on one x86 CPU thread, with PyTorch or ONNX Runtime; no GPU needed.
Part of the INDIC TINY TTS MODELS collection: 12 tiny open models, one per language, for 11 Indian languages and Indian English, with the datasets they were trained on.
Training data: sulabhkatiyar/indic-tts-mini-24k: the Kannada male voice (AI4Bharat Rasa, derived from the full-resolution sulabhkatiyar/indic-tts-multilingual-prepared) and the Kannada female voice (LIMMITS / SYSPIN corpus), 24 kHz.
Other open Kannada models, same test
FLORES-200 news sentences (100)
Fewer recognition errors than these 5 open Kannada models; the same as 2 larger models.
| Model | Size | CER, % | Extra errors vs this model, points (95 % CI) | Result |
|---|---|---|---|---|
| This model | 22.5 M | 1.05 | ||
ai4bharat/IndicF5 |
350.6 M | 0.70 | -0.36 [-0.55, -0.17] | theirs better |
Sammy-Mohan/Kokoro-82M-Kannada |
131.4 M | 0.96 | -0.09 [-0.31, +0.14] | same |
github.com/AI4Bharat/Indic-TTS (kn.zip) |
65.6 M | 1.23 | +0.18 [-0.08, +0.44] | same |
ai4bharat/vits_rasa_13 |
31.3 M | 1.78 | +0.73 [+0.42, +1.03] | ours better |
SYSPIN/vits_Kannada_Female + SYSPIN/vits_Kannada_Male |
29.1 M | 2.13 | +1.07 [+0.71, +1.45] | ours better |
ai4bharat/indic-parler-tts |
937.8 M | 2.30 | +1.25 [+0.39, +2.45] | ours better |
facebook/mms-tts-kan |
29.1 M | 3.31 | +2.26 [+1.73, +2.85] | ours better |
multilingual-tts/VITS-OpenBible-Kannada |
29.1 M | 5.90 | +4.85 [+4.01, +5.73] | ours better |
Held-out test sentences (299)
Fewer recognition errors than this open Kannada model; the same as one larger model.
| Model | Size | CER, % | Extra errors vs this model, points (95 % CI) | Result |
|---|---|---|---|---|
| This model | 22.5 M | 3.66 | ||
github.com/AI4Bharat/Indic-TTS (kn.zip) |
65.6 M | 3.84 | +0.18 [-0.69, +1.11] | same |
facebook/mms-tts-kan |
29.1 M | 9.42 | +5.76 [+4.72, +6.93] | ours better |
| Also trained on same speaker voices: | ||||
Sammy-Mohan/Kokoro-82M-Kannada |
131.4 M | 3.68 | +0.02 [-0.80, +1.03] | same |
ai4bharat/vits_rasa_13 |
31.3 M | 4.12 | +0.46 [-0.40, +1.26] | same |
SYSPIN/vits_Kannada_Female + SYSPIN/vits_Kannada_Male |
29.1 M | 6.16 | +2.49 [+1.33, +4.12] | ours better |
Larger models, same held-out sentences
- Same error rate as
ai4bharat/indic-parler-ttsat 1/42 the size, ~120× faster on T4 - Same error rate as
ai4bharat/IndicF5at 1/16 the size, ~127× faster on T4
| Model | Size | Sentences | CER, % | This model, same sentences, % | Extra errors vs this model, points (95 % CI) | Result |
|---|---|---|---|---|---|---|
| Also trained on same speaker voices: | ||||||
ai4bharat/IndicF5 |
350.6 M | 40 | 2.44 | 3.71 | -1.27 [-3.06, +0.76] | same |
ai4bharat/indic-parler-tts |
937.8 M | 40 | 3.27 | 3.71 | -0.44 [-2.08, +0.99] | same |
CER = character error rate, lower is better; size = parameters. Method: see the Appendix.
Speed
| Hardware | Setting | Faster than real time | Real-time factor |
|---|---|---|---|
| x86 CPU (Kaggle notebook, Intel Xeon @ 2.20GHz, 4 vCPU) | PyTorch, 1 thread | 7× | 0.136 |
| Apple M1 laptop, CPU | PyTorch, 1 thread | 38× | 0.026 |
| NVIDIA T4 GPU | PyTorch | 62× | 0.016 |
Real-time factor = total synthesis time ÷ total audio length over 30 held-out sentences (lower is faster).
Listen
Clips made by this model for sentences from the held-out evaluation set; never used in training.
Female voice (voice="female")
ಹಾಸಿಗೆ ಸುತ್ತಿ ಎತ್ತಿದಾಗ ದಿಂಬಿನಲ್ಲಿದ್ದ ಚೋಪಡಿ ಕೆಳಗೆ ಬಿತ್ತು.
Transliteration: Hasige sutti ettidaga dimbinallidda chopadi kelage bittu.
Translation (approx.): When the bedding was rolled up and lifted, the notebook kept in the pillow fell down.
ಇನ್ನೊಂದು ಅರ್ಥದಲ್ಲಿ ನಮ್ಮೊಳಗೆ ಆ ಸೂರ್ಯನಿದ್ದು ನಮ್ಮ ಬದುಕಿನ ಹತೋಟಿ ಅವನ ಕೈಯಲ್ಲಿದೆ ಎನ್ನಬಹುದು.
Transliteration: Innondu arthadalli nammolage a suryaniddu namma badukina hatoti avana kaiyallide ennabahudu.
Translation (approx.): In another sense, one could say that the sun is within us and that the control of our life is in his hands.
ಆಗ ಅಸಫ್ ಜಾ ತನ್ನ ನಂಬಿಕಸ್ಥ ವ್ಯಕ್ತಿಯಾದ ಅನ್ವರುದ್ದೀನನನ್ನು ಕಾರ್ನಾಟಿಕದ ನವಾಬನನ್ನಾಗಿ ನೇಮಿಸಿದನು.
Transliteration: Aga asaph ja tanna nambikastha vyaktiyada anvaruddinanannu karnatikada navabanannagi nemisidanu.
Translation (approx.): Then Asaf Jah appointed his trusted man Anwaruddin as the Nawab of the Carnatic.
ಅಂಗಳದಲ್ಲಿ ಬೆಂಕಿ ಮಾಡಿ ಸೌದೆ ಹಾಕ್ತಾ ಬೆಂಕಿ ಕಾಯಿಸ್ತಾ ಸುತ್ತ ಕುಳಿತೆವು.
Transliteration: Angaladalli benki madi saude hakta benki kayista sutta kulitevu.
Translation (approx.): We lit a fire in the courtyard and sat around it, adding firewood and warming ourselves.
Male voice (voice="male")
ಪ್ರತಿ ಸೈಟ್ಗೆ ಹೋಗಿ ಮತ್ತು ನಿಮ್ಮ ವ್ಯಾಪಾರ ಇಮೇಲ್ ಅನ್ನು ಬಳಸಿಕೊಂಡು ಖಾತೆಗೆ ಸೈನ್ ಅಪ್ ಮಾಡಿ.
Transliteration: Prati saitge hogi mattu nimma vyapara imel annu balasikondu khatege sain ap madi.
Translation (approx.): Go to each site and sign up for an account with your business email.
ಇದು ಹೋಳಿ ಹಬ್ಬ ಆದರೆ ಮಣಿಪುರದ ಜನರು ಯೌಸಂಗ್ ಎಂಬ ಪ್ರಾದೇಶಿಕ ಹೆಸರನ್ನು ನೀಡಿದ್ದಾರೆ.
Transliteration: Idu holi habba adare manipurada janaru yausang emba pradeshika hesarannu nididdare.
Translation (approx.): It is the Holi festival, but the people of Manipur have given it a regional name, Yaosang.
ಬಂಗಾಳದಲ್ಲಿ ಅತಿ ವಿಜೃಂಭಣೆಯಿಂದ ದುರ್ಗಾಪೂಜೆ ಎಂಬ ಹೆಸರಿನಲ್ಲಿ ಆಚರಿಸುವ ಈ ಹಬ್ಬ ತಾಯಿ ದುರ್ಗೆ ತೌರುಮನೆಗೆ ಮರಳುವುದರ ಸೂಚಕವಾಗಿದೆ.
Transliteration: Bangaladalli ati vijrimbhaneyinda durgapuje emba hesarinalli acharisuva i habba tayi durge taurumanege maraluvudara suchakavagide.
Translation (approx.): This festival, celebrated in Bengal with great splendour as Durga Puja, marks mother Durga's return to her parents' home.
ಟೆಸ್ಟ್ ಡ್ರೈವ್ ಮಾಡುವಾಗ, ಕಾರಿನ ಸಮಸ್ಯೆಗಳನ್ನು ಗಮನಿಸಿ.
Transliteration: Test draiv maduvaga, karina samasyegalannu gamanisi.
Translation (approx.): During the test drive, look out for any problems with the car.
If the players do not show, the files are in samples/.
Quick start
pip install torch numpy safetensors huggingface_hub
import sys
from huggingface_hub import snapshot_download
sys.path.insert(0, snapshot_download("sulabhkatiyar/indic-tiny-tts-kannada", ignore_patterns=["onnx/*"])) # model + its small Python package
from indic_tiny_tts import TinyTTS
tts = TinyTTS.from_pretrained("sulabhkatiyar/indic-tiny-tts-kannada") # uses the NVIDIA GPU if there is one, else the CPU
wav = tts.synthesize("ನಮಸ್ಕಾರ! ಇಂದು ಹವಾಮಾನ ತುಂಬಾ ಚೆನ್ನಾಗಿದೆ.", voice="female")
tts.save_wav(wav, "hello.wav")
wav is a NumPy array of 24 kHz audio. voice is "female" or "male". Optional: speed=1.2 speaks faster, and seed=0 gives the same output every time.
Choose where it runs
tts = TinyTTS.from_pretrained("sulabhkatiyar/indic-tiny-tts-kannada", device="cpu") # any computer
tts = TinyTTS.from_pretrained("sulabhkatiyar/indic-tiny-tts-kannada", device="cuda") # NVIDIA GPU
tts = TinyTTS.from_pretrained("sulabhkatiyar/indic-tiny-tts-kannada", device="cpu", threads=1) # Mac with Apple silicon (fastest there)
On a Mac, device="mps" (the Apple GPU) also works, but for a model this small the CPU with one thread is faster.
Without PyTorch (ONNX Runtime)
The onnx/ folder holds the same model as three ONNX graphs. This path needs only NumPy and ONNX Runtime:
pip install onnxruntime numpy huggingface_hub
import sys
from huggingface_hub import snapshot_download
folder = snapshot_download("sulabhkatiyar/indic-tiny-tts-kannada", ignore_patterns=["*.safetensors"])
sys.path.insert(0, folder)
from indic_tiny_tts import TinyTTSOnnx
tts = TinyTTSOnnx.from_pretrained(folder)
wav = tts.synthesize("ನಮಸ್ಕಾರ! ಇಂದು ಹವಾಮಾನ ತುಂಬಾ ಚೆನ್ನಾಗಿದೆ.", voice="male")
tts.save_wav(wav, "hello.wav")
Install as a package (optional)
This adds an indic-tiny-tts command:
hf download sulabhkatiyar/indic-tiny-tts-kannada --local-dir indic-tiny-tts-kannada
pip install ./indic-tiny-tts-kannada
indic-tiny-tts "ನಮಸ್ಕಾರ! ಇಂದು ಹವಾಮಾನ ತುಂಬಾ ಚೆನ್ನಾಗಿದೆ." --model ./indic-tiny-tts-kannada --voice male -o hello.wav
Measure the speed
Inside the downloaded folder:
python -m indic_tiny_tts.latency --model . --device cpu --threads 1
python -m indic_tiny_tts.latency --model . --backend onnx --threads 4 # the ONNX version
Details
| Language | Kannada (Kannada script) |
| Voices | female (LIMMITS 2024, IISc SYSPIN), male (Rasa, AI4Bharat) |
| Audio | 24 kHz, mono |
| Size | 22.5 million parameters (about 90 MB per format: PyTorch or ONNX) |
| Runs on | CPU (Windows, Linux, Mac), NVIDIA GPU (CUDA), Apple silicon |
| Needs | Python 3.9+ and PyTorch, or ONNX Runtime. No eSpeak or other system packages. |
How good is it?
On held-out sentences a speech recogniser misheard 3.66 % of this model's characters, close to the 2.49 % of the real recordings.
Good to know
- Kannada script only. English words in Latin letters (
hello) are skipped with a warning. Write them in Kannada script. - Whole numbers are read out in Kannada:
25becomes ಇಪ್ಪತ್ತೈದು and2024becomes ಎರಡು ಸಾವಿರದ ಇಪ್ಪತ್ನಾಲ್ಕು. Decimals, percentages and currency (3.5%,₹500) are not converted yet and are skipped, so write them in words. Times such as10:30and phone numbers are not read correctly yet, so write them out in words if that matters. - Long text is fine. It is split into sentences at
।,?,!and., which are spoken one after another. - Style: a calm reading voice. The model learned from about 36 hours of recorded speech by two speakers.
- The male voice is the weaker one. It makes more recognition errors (5.44 % against 1.90 % for the female voice; its own recordings score 3.53 %).
- The female voice can hold a sound a little long now and then. Our automatic check for drawn-out sounds flagged it.
- Speaker consent: the female voice comes from the LIMMITS 2024 corpus. Speaker consent for voice-model release is not documented by the source; the corpus is released under CC-BY-4.0 by its organisers.
- Speaking rate: each voice has a default speed matched to its speaker on held-out data (
length_scaleinconfig.json: kn_M 0.964, kn_lim_F 1.032).speed=scales it further. - Please use it responsibly. Do not use it to pass yourself off as the original speakers or to mislead people.
Training data and credits
- Rasa dataset by AI4Bharat (ai4bharat/Rasa), licence CC-BY-4.0: the Kannada male speaker (17.4 h), the read, conversation and command recordings, leaving out the emotional ones. Our changes: audio resampled from 48 kHz to 24 kHz; clips shorter than 0.5 s or longer than 20 s left out; text normalised (for example, numbers written out in words); training clips that a speech recogniser could not match to their text (character error rate above 25 %, or a length mismatch) left out: 1.29 % of the male clips.
- LIMMITS 2024 challenge corpus (SYSPIN project, © Indian Institute of Science, Bangalore, SPIRE Lab), licence CC-BY-4.0 through the SYSPIN Dataset Terms of Use, obtained from the Hugging Face mirror iAkashPaul/limmits-2024 (revision
9ee2415a): the Kannada female speaker (18.5 h used). Our changes: audio resampled to 24 kHz mono; written/spoken transcript pairs resolved to the spoken form; text normalised; clips shorter than 0.5 s or longer than 20 s and exact duplicates left out; held-out sentences removed from training; training capped at 20 h; leading silence longer than 0.3 s shortened to 0.15 s (6,387 of 6,789 clips); training clips failing the speech recogniser check left out: 0.00 %. The organisers note that the release contains some audio and text errors. - Prepared data: sulabhkatiyar/indic-tts-mini-24k (revision
8fe16c93). - Vocoder: Vocos, with the pretrained charactr/vocos-mel-24khz weights used unchanged (MIT licence; trained by its authors on LibriTTS).
- Model design: based on Matcha-TTS (MIT licence). The acoustic model (8.9 million parameters) reads characters directly (no phonemiser) and was trained from scratch on the data above for 120,000 steps (batch 32, about 10 hours on one NVIDIA T4); the released weights are the averaged (EMA) weights at that step.
- Number reading: word tables adapted from AI4Bharat's indic-numtowords (MIT licence).
- Evaluation tools: AI4Bharat IndicConformer (speech recognition) and SpeechBrain ECAPA (speaker similarity).
Licence
- Acoustic model weights (
model.safetensors,onnx/text_encoder.onnx,onnx/flow_decoder.onnx): CC-BY-4.0. You may use them, also commercially, if you give credit to this model and to the training data (the Rasa dataset and the LIMMITS 2024 corpus; see Citation and LICENSE.md). - Vocoder weights (
vocoder.safetensors,onnx/vocoder.onnx): the unchanged pretrained Vocos weights, MIT licence by their authors. - Code (
indic_tiny_tts/): MIT. - Details and third-party notices: LICENSE.md.
Citation
If you use this model, please cite it:
@misc{katiyar2026indictinyttskannada,
author = {Sulabh Katiyar},
title = {Indic Tiny TTS: Kannada},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/sulabhkatiyar/indic-tiny-tts-kannada}}
}
Please also cite the data and the work it builds on (for the LIMMITS 2024 voice, credit the SYSPIN project and IISc SPIRE Lab as in its terms of use):
@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}},
booktitle = {Proc. INTERSPEECH 2024},
year = {2024}
}
@inproceedings{mehta2024matcha,
author = {Shivam Mehta and Ruibo Tu and Jonas Beskow and {\'E}va Sz{\'e}kely and Gustav Eje Henter},
title = {{Matcha-TTS}: A fast {TTS} architecture with conditional flow matching},
booktitle = {Proc. ICASSP 2024},
year = {2024}
}
@inproceedings{siuzdak2024vocos,
author = {Hubert Siuzdak},
title = {Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis},
booktitle = {Proc. ICLR 2024},
year = {2024}
}
Appendix: how we evaluated
Every model spoke the same text: 299 held-out test sentences from both voices, never used in training, and 100 FLORES-200 news sentences. A speech recognition model (AI4Bharat IndicConformer-600M) transcribed each clip, and the score is its character error rate (CER). A paired bootstrap 95 % confidence interval of the CER difference over the same sentences decides each result: better when the whole interval favours this model, the same when it includes zero, worse when it favours the other model. The larger models read the first 40 of the held-out sentences (20 per voice) unless the table shows a larger number, and all 100 FLORES-200 sentences.
- Downloads last month
- 135


