Mithi v1 Flash

Mithi v1 Flash

Reads text aloud in 11 Indian languages with a male and a female voice, on a GPU or a plain CPU. Made for developers who add speech to apps.

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 v1 Flash is a small text-to-speech model from Oomni that speaks 11 Indian languages, English among them. It has two voices, Satvik (male) and Charvi (female), and writes 24 kHz mono audio. The model has 99.1 M parameters and takes 197 MB.

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

Industry | Who uses it | How they use it | Example
Customer service | Support teams | Read replies aloud in the caller's own language | An automated phone message in Tamil or Hindi
Education | Teachers and app makers | Add spoken lessons in regional languages | A reading app that speaks Marathi stories
Media | Content creators | Narrate short news or explainer text | A news summary voiced in Bengali
Accessibility | App developers | Let on-screen text be heard | A screen reader add-on for Telugu text

If you only need English speech or want many voices to choose from, you can skip it.

Where to use it:

• Windows, macOS and Linux machines with Python
• A command line, through the included inference script
• Your own Python code, through the Mithi class
• CPU-only devices such as laptops, a Raspberry Pi, phones and servers without a GPU

Use case | Who | What they get back
Voice messages | Developers | A WAV file spoken in the language and voice you pick
Reading long text | App makers | Audio that follows the text one sentence at a time
Numbers and times | Content teams | Digits, times and currency spoken out as words

How to set it up: You need these first:

• Python 3.10 or newer, with pip
• A GPU is optional, since the model also runs on a CPU
• About 197 MB of disk space for the weights

1. Install the Hugging Face hub tool and download the model into a local folder.
2. Move into that folder.
3. Install the Python requirements.
4. Download the English spaCy model named in the last command.

pip install -U huggingface_hub
hf download Oomni/mithi-v1-flash --local-dir mithi-v1-flash
cd mithi-v1-flash
pip install -r requirements.txt
python -m spacy download en_core_web_sm

How to use it: Run the inference script from inside the model folder and give it a language code, a voice, your text and an output file name. The example below speaks one English line in the male voice. Add the CPU option when you have no GPU.

python inference.py --lang en --voice satvik --text "Your appointment is confirmed for Monday at 10:30 AM." --out english.wav

Tool | What it does
--lang | Picks the language by its short code, such as hi, en or ta
--voice | Picks satvik or charvi
--text | Holds the text to read aloud
--out | Names the WAV file to write
--device cpu | Runs the model on the CPU
--threads | Sets how many CPU cores to use

Example: You run the command above with the line about a confirmed appointment. The script reads it in the Satvik voice, spells out the time as words, and saves english.wav as 24 kHz mono audio that you can play.

For newcomers and experts: Mithi turns typed text into spoken audio. A good first step is to download the folder, run the English command with the satvik voice, and play the file it writes.

Under the hood it is a bf16 model built from open-source parts such as Kokoro-82M and indic-bert, plus a text normaliser. It trades speed for size on a CPU: a 15-word sentence takes about 3 s with 4 threads on a server chip, and slower chips take longer.

Text and language code go in
The normaliser spells out numbers
The model speaks in the chosen voice
A 24 kHz mono WAV file comes out

Good to know: Speed was measured on an AMD EPYC 9554 server CPU for a sentence of about 7 s of speech.

Title: Seconds per 15-word sentence on a server CPU
1 thread: 8 s
4 threads: 3 s

• Languages: Hindi, English, Bengali, Gujarati and Kannada, plus Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu
• Long text: it is read sentence by sentence automatically
• Parts: built on permissive open-source models and tools, and packages in requirements.txt keep their own licenses
• License: Apache 2.0, free for personal, research and commercial use, with the license and credits kept with any copy you share
• Cost to run: none named, you use your own GPU or CPU
• Data it reaches: the text you give it, and the audio files it writes where you point it

On Hugging Face, 13 November 2026

• Task: text to speech
• Formats: Safetensors
• Downloads in the last 30 days: 592
• Likes: 70
• 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-v1-flash
• Files: https://huggingface.co/Oomni/mithi-v1-flash/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Oomni/mithi-v1-flash

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