Pratilekha
On-device STT models with Indic Language Support (8). Focused on Devnagari Languages in noisy environments, these on-device STT models are specifically tuned for different accents, and real world applications and not just for crushing benchmarks in lab tests.
Speech To TextSTTS2THuggingFacePyTorchMLX
The linked HF repos are actually the v0 checkpoints of the Pratilekha family of models by Alchemyst. We’ve focused on Hindi and Bengali alongside English because those are the languages I understand XDD.
Also, it has been optimized for noisy environments. Vistaar style benchmarks are something I’m working on, will be out in sometime too.
Language switching needs a lot of work, as with all whisper based models, trying to find workarounds for them 😅