Original title: Oomni/mithi-v1-flash
A small text-to-speech model released by Oomni, supporting 11 languages: Hindi, English, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu; two voices, Satvik (male) and Charvi (female). The model has 99.1M parameters and is 197 MB in size, outputs 24 kHz mono audio, and is built on open-source components such as Kokoro-82M, Maya2-native, and indic-bert. The authors state that on a server CPU (AMD EPYC 9554), generating approximately 7 seconds of speech takes about 8 seconds with a single thread and about 3 seconds with 4 threads.
Why it matters Under 200 MB in size, usable without a GPU, Apache-licensed weights allow commercial use, providing a lightweight starting point for Indian language speech applications
Rising score 60 · Likes 64 · Downloads 436
Reasons
- Use a GPU if available; it also runs on CPU. The authors state it can run on laptops, Raspberry Pi, and mobile phones.
- Includes a text normalizer that automatically expands numbers, times, and currencies into words; long texts are automatically read sentence by sentence.
- Weights are licensed under Apache-2.0, allowing commercial use.
Usage
- Run
hf download Oomni/mithi-v1-flash --local-dir mithi-v1-flash, enter the directory, runpip install -r requirements.txt, and download the spaCy English model. - Run
python inference.py --lang hi --voice charvi --text "…" --out hindi.wav; for CPU execution, add--device cpu --threads 4.
Risks
- The authors state there are no public evaluations yet, with neither automatic metrics nor human listening scores.
- Intonation and emotion are flat; speech rate is too fast for Kannada, Malayalam, Marathi, and Odia; rare names and loanwords may be mispronounced.
- Chinese is not supported; this entry has not been tested, so effectiveness should be judged based on the model card and sample audio.