Forced alignment, which maps audio to text transcripts, is a common step in speech research and downstream applications. A new preprint on arXiv argues that these aligners perform poorly on spontaneous, noisy real-world speech, a setting far from the clean read speech they are often trained on.
The authors identify fillers, laughter, and disfluencies as major sources of word-boundary corruption, which can propagate errors into tasks like phonetic analysis or language learning tools. They note that this degradation impacts many downstream applications, though the abstract does not quantify the extent.
To address the problem, the researchers revisit the classical GMM-HMM aligner, a traditional approach that may offer more robustness to acoustic variability. The abstract does not include experimental results, so it remains unclear whether this older method outperforms modern neural aligners in practice.