Phonetic, tonal, and speaker information occupy largely orthogonal subspaces across pretrained wav2vec2 models of four languages
measured in 1 paper- Pretrained 12-layer wav2vec 2.0 base checkpoints (English wav2vec2-base, French wav2vec2-FR-7K-base, Mandarin mandarin-wav2vec2, Vietnamese wav2vec2-base-vi) are analyzed on test speech in English, French, Mandarin, Vietnamese and Thai (no Thai-pretrained model); subspaces for phones, lexical tones and speakers are built by PCA on per-class centroids of frame representations and their pairwise orthogonality is measured with Cumulative Residual Variance (CRV). [gubian-etal-2025-pretraining-language-phonetic-tonal-speaker-info-ssl-speech] - The authors report that "the subspaces encoding phones, tones, and speakers are largely orthogonal" for all pretraining and test languages; phone/speaker subspaces are the most orthogonal, tone/speaker and tone/phone slightly less so. [gubian-etal-2025-pretraining-language-phonetic-tonal-speaker-info-ssl-speech] - Orthogonality shows no obvious relation to pretraining language for phone/speaker; Tone-vs-Phone orthogonality varies mainly with the test language (Mandarin > Vietnamese > Thai), ordered like the tone-phone statistical dependence in each language's test corpus. Linear probes (phones, tones, speakers) show a modest matched-language advantage for phones and tones in middle layers. [gubian-etal-2025-pretraining-language-phonetic-tonal-speaker-info-ssl-speech]