jonatasgrosman/huggingsound

Provides unified APIs for speech recognition inference, fine-tuning, and evaluation using Hugging Face Hub models with optional language model decoding via KenLM or Flashlight. Supports character-level timestamp and confidence outputs, multiple LM decoder backends (Kensho, Parlance, Flashlight), and handles various audio formats through ffmpeg integration. Built on PyTorch and Transformers, targeting wav2vec2 and similar CTC-based models for multilingual speech tasks.

470 stars and 404 monthly downloads. No commits in the last 6 months. Available on PyPI.