shashikg/WhisperS2T

Supports multiple inference backends (Original OpenAI, HuggingFace with FlashAttention2, CTranslate2, TensorRT-LLM) with intelligent audio batching, VAD integration, and dynamic time-length processing to reduce computation overhead. Incorporates hallucination-reduction heuristics and asynchronous large-file loading while simultaneously transcribing batched segments, enabling multi-language and multi-task decoding in a single batch pass.

541 stars. No commits in the last 6 months.