SPAM uses Phonological Activation Mapping for phone segmentation and recognition

Researchers propose SPAM, a method that leverages self-supervised speech models (S3Ms) to simultaneously solve phone segmentation and recognition by steering latent phonetic structure. The approach maps each S3M representation frame to a vector of phonological feature activations, such as voicing and nasality.

• SPAM utilizes lightweight, gradient-descent-free prediction heads for both recognition and segmentation tasks.
• The method requires less than a minute of phonetic transcriptions for training.
• It generalizes to unseen phones during training and achieves strong performance across diverse datasets.