Mithi v2 Flash

A small text-to-speech model for 11 Indian languages, with 10 voices: 5 male and 5 female.

Languages: Hindi, English, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu.
Voices: Satvik, Anirban, Nakul, Vasant, Siddhant (male); Charvi, Ahana, Vaidehi, Malini, Jahnavi (female).
Size: 99.1 M parameters, 199 MB.
Weights: bf16 safetensors.
Output: 24 kHz mono audio.
License: Apache-2.0, free to use, including commercially. See License.

Demo

All voices in the video are generated by Mithi v2 Flash.

License

Mithi is released under the Apache License 2.0 (see LICENSE). It is free to use, for personal, research and commercial projects alike: you can use, modify and redistribute the model and code, and build products on it. Keep the license and the credits (see Acknowledgements) with any copy you share.

Mithi was built entirely on open-source models and tools, each under a permissive license:

Component	License
Maya2-native	Apache-2.0
Kokoro-82M	Apache-2.0
indic-bert	MIT
Maya text normaliser	Apache-2.0
misaki	Apache-2.0

Packages installed with pip keep their own licenses.

Quick start

1. Download the model and install the dependencies (Python 3.10 or newer):

pip install -U huggingface_hub
hf download Oomni/mithi-v2-flash --local-dir mithi-v2-flash
cd mithi-v2-flash
pip install "torch>=2.4" "transformers>=5.0" "safetensors>=0.4" "sentencepiece>=0.2" "numpy>=1.26" "soundfile>=0.12" \
    "misaki[en]>=0.9.4" "espeakng-loader>=0.2.4" "phonemizer-fork>=3.3.2" "spacy>=3.8" "num2words>=0.5"
python -m spacy download en_core_web_sm

espeak-ng comes with the espeakng-loader package, so no system install is needed. A GPU is used if there is one; the model also runs on CPU.

2. Run it from the command line:

python run.py --lang hi --voice ahana --text "आज मौसम कितना अच्छा है! चलो, शाम को कहीं बाहर घूमने चलते हैं।" --out hindi.wav
python run.py --lang en --voice anirban --text "Your appointment is confirmed for Monday at 10:30 AM." --out english.wav

3. Or from Python, inside the mithi-v2-flash folder:

from run import Mithi

tts = Mithi()                                      # loads the model from this folder
audio = tts.synthesize("नमस्ते! आप कैसे हैं?", lang="hi", voice="vaidehi")
tts.save(audio, "namaste.wav")                     # 24 kHz WAV

for voice in tts.voices:                           # all 10 voices
    tts.save(tts.synthesize("வணக்கம்! எப்படி இருக்கீங்க?", lang="ta", voice=voice), f"hello_ta_{voice}.wav")

4. Run on CPU (laptops, Raspberry Pi, phones, servers without a GPU):

python run.py --device cpu --threads 4 --lang hi --voice charvi --text "आज मौसम कितना अच्छा है! चलो, शाम को कहीं बाहर घूमने चलते हैं।" --out hindi_cpu.wav

--threads sets how many CPU cores to use; without it PyTorch uses all of them. From Python:

import torch
torch.set_num_threads(4)

from run import Mithi
tts = Mithi(device="cpu")
tts.save(tts.synthesize("नमस्ते! आप कैसे हैं?", lang="hi", voice="nakul"), "namaste_cpu.wav")

Measured speed for an average 15-word sentence (about 7 s of speech), on a server CPU (AMD EPYC 9554):

CPU threads	Time per sentence
1	~8 s
4	~3 s

Slower CPUs, such as a Raspberry Pi, take proportionally longer.

Languages: hi Hindi · en English · bn Bengali · gu Gujarati · kn Kannada · ml Malayalam · mr Marathi · or Odia · pa Punjabi · ta Tamil · te Telugu

Long text is read sentence by sentence automatically.

Voices

Ten voices, five male and five female. Every voice speaks all 11 languages.

Voice	`--voice`	Gender	Since
Satvik	`satvik`	Male	v1
Charvi	`charvi`	Female	v1
Ahana	`ahana`	Female	v2
Vaidehi	`vaidehi`	Female	v2
Malini	`malini`	Female	v2
Jahnavi	`jahnavi`	Female	v2
Anirban	`anirban`	Male	v2
Nakul	`nakul`	Male	v2
Vasant	`vasant`	Male	v2
Siddhant	`siddhant`	Male	v2

Files

File	Contents
`run.py`	everything needed to run the model: command line and Python API
`config.json`	languages, voices and settings
`*.safetensors`, `indic_encoder/`	model weights
`LICENSE`	Apache License 2.0
`assets/`	demo video and its thumbnail

Limitations

Expressiveness: intonation and emotion are flatter than in natural human speech.
Kannada, Malayalam, Marathi, Odia: speech can come out a little faster than natural.
Punjabi with Jahnavi or Vasant may sound less natural than with the other voices.
Pronunciation: some words, especially rare names and loanwords, can be mispronounced.
Numbers, dates, currency, URLs are read only as well as the text normaliser expands them.
Long text is read sentence by sentence; each sentence is spoken without seeing the others.
No published evaluation yet: no automated scores and no human listening study (MOS).

Acknowledgements

Maya2-native by Maya Research, Apache-2.0.
Kokoro-82M: hexgrad/kokoro, Apache-2.0.
indic-bert: ai4bharat/indic-bert, MIT.
misaki and espeak-ng.