GPU forced alignment

easyaligner has been released as a GPU-accelerated forced alignment library.

• It automatically detects the relevant segments even when the transcript does not cover the entire audio.
• It is designed to handle segments that include unnecessary speech at the start/end.
• It states that long audio/text can be aligned at once without chunking.
• It aligns using normalized text, then maintains a mapping so the original formatting can be restored.
• Even when splitting the result into desired granularities such as sentence or paragraph, the original formatting is preserved.
• The backend uses the PyTorch forced alignment API and a GPU-based Viterbi implementation.
• It is compatible with wav2vec2-family models, and the documentation also provides tutorials on various alignment scenarios and custom text processing.