Mithi v2 Flash

Mithi v2 Flash

Reads text aloud in 11 Indian languages, English included, with 10 voices, from a 199 MB model that runs on a GPU or a plain CPU. For developers who need small, free text-to-speech.

Made by Oomni on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Mithi v2 Flash is a small text-to-speech model with 10 voices, five male and five female. It speaks Tamil, Telugu, Kannada, Malayalam, Hindi, Bengali, Gujarati, Marathi, Odia, Punjabi and English. The model has 99.1M parameters, weighs 199 MB and writes 24 kHz mono audio.

Who it suits: It suits developers and teams who need spoken output in Indian languages without a large model or a GPU.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Customer service | Support teams | Read confirmations aloud in a customer's language | A booking confirmation spoken in English or Hindi
Education | Teachers and course makers | Narrate lessons in regional languages | A short lesson read aloud in Tamil
Accessibility | App developers | Read on-screen text aloud | A Marathi article spoken by one of the voices
Edge devices | Hardware tinkerers | Run speech on a laptop or a small board | A Raspberry Pi reading a sentence from a CPU

People who need highly expressive or emotional speech can skip it, since the card says intonation and emotion are flatter than natural speech.

Where to use it:

• The command line: run.py takes a language, a voice, a text and an output file
• Python code: a class called Mithi turns text into audio
• A GPU or a CPU: it uses a GPU when present and also runs on laptops, Raspberry Pi, phones and servers without one
• Windows, macOS or Linux: needs Python 3.10 or newer

Title: Jobs it does
Use case | Who | What they get back
Voice-over for apps | Developers | A WAV file of the text in the chosen voice
Spoken notifications | Service teams | Short spoken messages in an Indian language
Offline speech on small hardware | Tinkerers | Audio generated on a CPU with no GPU

How to set it up: You need these first:

• Python 3.10 or newer on Windows, macOS or Linux
• About 199 MB of storage for the weights
• The Python packages listed in the commands below

1. Install the Hugging Face hub package.
2. Download the model into a folder named mithi-v2-flash.
3. Move into that folder.
4. Install the packages with pip, then download the spaCy English model.

You do not need a separate system install of espeak-ng, because the espeakng-loader package brings it along.

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

How to use it: Run run.py with a language code, a voice and your text, and it writes a WAV file. These two commands speak one Hindi and one English sentence.

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

Title: Command line options
Tool | What it does
run.py | Runs the model from the command line and offers a Python API
--lang | Picks the language by its short code, such as hi or en
--voice | Picks one of the ten voices
--text | Gives the text to speak
--out | Sets the WAV file to write
--device cpu | Runs on the CPU instead of a GPU
--threads | Sets how many CPU cores to use

From Python, run this 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")

To run on the CPU, add the device flag. Without the threads flag, PyTorch uses every core.

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

Longer text is split into sentences and spoken one after another without extra steps. Every voice speaks all 11 languages. The male voices are Nakul, Siddhant, Vasant, Anirban and Satvik, and the female voices are Ahana, Charvi, Jahnavi, Malini and Vaidehi.

Example: You run the command for English with the voice Nakul and ask it to say "Your table is booked for Friday at 7 PM". A 24 kHz mono WAV file comes back with that sentence spoken.

For newcomers and experts: If you are new to this, treat it as a tool that turns typed text into a spoken audio file. Run the English command once, listen to the WAV and then change the voice name.

If you work in the field, the model is built from open-source parts, among them Maya2-native, Kokoro-82M, indic-bert, a Maya text normaliser and misaki, all under permissive licenses. The weights are bf16 safetensors. Odia, Marathi, Malayalam and Kannada speech may run slightly quicker than a natural speaker. Two voices date from v1 and eight from v2.

Pick a language code and a voice
Give it the text to speak
Long text is split into sentences
A 24 kHz mono WAV file is written

Good to know: The card timed an average sentence of 15 words, roughly 7 seconds of audio, on a server CPU. Slower CPUs take proportionally longer.

Title: Seconds to speak one sentence on a server CPU
1 CPU thread: 8 seconds
4 CPU threads: 3 seconds

• License: Apache-2.0, so personal, research and commercial use are all allowed, and any copy you share must carry the license and credits
• Cost to run: none named
• Data it reaches: the text you give it and the WAV files it writes, while setup downloads Python packages and a spaCy English model
• Not for: highly expressive speech, since intonation and emotion are flatter than a natural voice

On Hugging Face, 13 November 2026

• Task: text to speech
• Formats: Safetensors
• Downloads in the last 30 days: 62
• Likes: 6
• License: apache-2.0
• Languages: hi, en, bn, gu, kn, ml, mr, or
• Tags: safetensors, text-to-speech, indic, bf16

Links

• Model: https://huggingface.co/Oomni/mithi-v2-flash
• Files: https://huggingface.co/Oomni/mithi-v2-flash/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Oomni/mithi-v2-flash

Then follow the usage in its model card. Nothing in this kit runs by itself.