Open-Speech-EkStep/vakyansh-models

Built on wav2vec2 and Conformer architectures, the models employ self-supervised pretraining on 34,000+ hours of unlabeled Indic audio followed by language-specific fine-tuning, with cross-lingual representations (CLSRIL-23) enabling transfer across 23 languages. Beyond ASR, the toolkit includes complementary models for punctuation restoration, text-to-speech synthesis, language identification, and gender classification. TorchScript export formats enable deployment across edge and production environments without framework dependencies.

325 stars. No commits in the last 6 months.