Fast Forced Aligner

Fast Forced Aligner is an open-source Python package by Bryan Thornbury (brthor) that performs forced alignment from existing CTC model outputs on NVIDIA GPUs. It takes log-probability emissions and transcript token IDs, then returns a token label and score for every audio frame. The package is pre-release at version 0.1.0 and is installed directly from GitHub.

What It Is

The library handles the alignment step only: acoustic inference stays with the caller's own CTC model, and the examples use stock Hugging Face Transformers models such as facebook/wav2vec2-base-960h to produce emissions. Recordings of different lengths are batched without padding, and results stay on the GPU as CUDA tensors. Two modes sit behind the same Python API: an optimized kernel (the default) and a conventional full-record CTC Viterbi DP kernel.

How the alignment modes work

Optimized mode solves relaxed 128-frame chunks in parallel, uses greedy shortcuts, repairs chunk boundaries, and falls back to full-record DP when needed. Conventional mode runs the full recurrence with specialized CUDA kernels and parallel traceback. The README notes that optimized mode is not universally bitwise equivalent to conventional DP, that adversarial near-tie inputs can differ, and that it can lose to conventional DP on some long recordings.

Inputs, outputs, and validation

Inputs are contiguous CUDA FP32 emissions and CUDA int64 target IDs on one device. Outputs keep blanks and repeats, and each result exposes paths, labels, scores, lengths, and a status code. Default GPU-side target validation avoids host waits, and an optional emission-validity scan can exclude invalid records. The docs also explain converting frame labels to timestamps using the model's frame stride, and include a comparison guide for users coming from TorchAudio's forced_align.

Setup path

It requires Linux, Python 3.11+, PyTorch 2.8.0, CUDA 12.8, an NVIDIA GPU and driver, and the CUDA toolkit with nvcc and a C++ compiler, because CUDA source is compiled on first use. A Docker example and a pinned benchmark subproject are included. The author reports benchmark results on an RTX 3090 against TorchAudio, k2, NeMo, and ctc-forced-aligner.