fast-forced-aligner
A CUDA library for batch aligning CTC frame scores and transcript tokens on a GPU.
What does it do?
When users supply log probabilities from their own CTC model and transcript token IDs, fast-forced-aligner returns a token label and score for every frame of each recording. It keeps inputs and results on the CUDA device and processes recordings of different lengths together. The same Python API offers a choice between an optimized alignment kernel and a conventional dynamic programming kernel that computes the full range. Acoustic model inference is outside the package’s scope, so users must first generate frame scores with their own model.
Key features
- Returns per-frame labels and scores from CTC frame scores and transcript token IDs.
- Batch processes recordings of different lengths on a GPU without padding.
- Selects an optimized alignment kernel or conventional CTC dynamic programming kernel through the same Python API.
- Keeps inputs and alignment results on the CUDA device.