Whistler Identification in Whistled Spanish (Silbo): A Case Study
Résumé
Deemed one of the world's most representative whistled languages, the Canary Islands' whistled Spanish, locally known as Silbo, has long attracted linguistic research. However, most studies have adopted linguistic, ethnological, or bioacoustic perspectives, overlooking the potential of computational methods within the digital humanities. This work advances the computational study of Silbo by presenting the first automated approach to Speaker Identification (SI)-i.e., the process of determining the speaker of a given utterance by computational meansin a closed-set configuration for this language. The proposal leverages standard feature extraction methods as well as pre-trained Speech Recognition models to extract representative embeddings and incorporates class-balancing mechanisms to mitigate biases arising from uneven representation of whistlers in the data-i.e., label imbalance. The results obtained on the only existing dataset specifically designed for computational analysis of Silbo, comparing three representative feature extraction methods, three oversampling policies, and five classification strategies, validate the proposal, achieving F1 scores close to 90% in the best-case scenarios. While laying a solid foundation for SI in Silbo, this study also highlights the scarcity of computational research on whistled languages, and particularly Silbo, emphasizing the need for further work to bridge traditional linguistic research and modern digital humanities.