Abstract

This study examines language variation bias in Indonesian Text-to-Speech (TTS) systems through a mixed-methods approach. Two research questions guide the investigation: (1) how accurately current TTS systems reproduce major regional accents, and (2) how users perceive the "neutrality" of TTS voices and which sociolinguistic factors influence these perceptions. The dataset comprised 1,500 TTS-generated samples and human recordings in four regional accents (Javanese, Sundanese, Batak, Minangkabau), alongside 1,500 user comments and survey responses from 250 participants.On the technical side, acoustic features were extracted and classified using a CNN model with stratified 5-fold cross-validation. Evaluation metrics included accuracy, precision, recall, F1-score, AUC, and visualization through confusion matrices and t-SNE clustering. Results showed higher accuracy for Javanese (F1 = 0.88) and Sundanese (F1 = 0.85), but weaker performance for Batak (F1 = 0.76) and Minangkabau (F1 = 0.73), revealing bias toward dominant accents. On the perceptual side, sentiment analysis with IndoBERT found 42% positive evaluations emphasizing clarity, while 23% were negative, citing Jakarta-accent bias, robotic quality, and lack of expressiveness. Topic modeling (LDA and BERT embeddings) identified six thematic clusters, including neutrality, accent bias, rhythm, and emotional resonance. Survey data confirmed that Jakarta-based voices were rated highest for neutrality and clarity but lowest for expressiveness, while regional accents were perceived as marked or informal. The novelty of this study lies in its integration of computational modeling with sociolinguistic analysis to examine Indonesian TTS for the first time. Unlike prior work focusing solely on technical performance, this research demonstrates that neutrality is not linguistically universal but socially constructed around Jakarta-based speech, providing new insights for developing more inclusive and culturally sensitive speech technologies.

Concepts :

Access to Document

SDGs

Citations by Year