OpenPA
A multi-granularity automatic pronunciation assessment framework that does not require forced alignment. It is based on the cross-attention mechanism of the Whisper model, supports both constrained and open response scenarios, and demonstrates zero-shot generalization capability across datasets.
Timeline (last 90 days)
OpenPA adopts a hybrid transcription adaptation technique, enabling the same model to handle heterogeneous transcribed text from both constrained-response and open-response settings
OpenPA adopts an adaptive substitution weighting technique that dynamically adjusts the penalty weight for pronunciation substitution errors based on the phonetic distance or frequency of the substitution
OpenPA is trained solely on the speechocean762 dataset
In the in-domain constrained-response scenario, OpenPA achieves the strongest overall performance among all participating systems
OpenPA achieves zero-shot performance retention on the unseen MultiPA dataset, maintaining competitive scoring capability without fine-tuning or annotation
Researchers conducted a cross-task evaluation of OpenPA on the speech of oral cancer patients, and preliminary results suggest its applicability may extend to medical scenarios such as pathological speech assessment
All Facts (6)
OpenPA adopts a hybrid transcription adaptation technique, enabling the same model to handle heterogeneous transcribed text from both constrained-response and open-response settings
50%UnverifiedOpenPA adopts an adaptive substitution weighting technique that dynamically adjusts the penalty weight for pronunciation substitution errors based on the phonetic distance or frequency of the substitution
50%UnverifiedOpenPA is trained solely on the speechocean762 dataset
50%UnverifiedIn the in-domain constrained-response scenario, OpenPA achieves the strongest overall performance among all participating systems
50%UnverifiedOpenPA achieves zero-shot performance retention on the unseen MultiPA dataset, maintaining competitive scoring capability without fine-tuning or annotation
50%UnverifiedResearchers conducted a cross-task evaluation of OpenPA on the speech of oral cancer patients, and preliminary results suggest its applicability may extend to medical scenarios such as pathological speech assessment
50%