PaperlayerView exact revision on arXiv

Source-linked AI summary

Scaling Self-Supervised Speech Models Uncovers Deep Linguistic Relationships: Evidence from the Pacific Cluster

Minu Kim, Hoirin Kim, David R. Mortensen

arXiv:2603.07238v2cs.CLeess.AS

TL;DR

Prior work found that S3M language representations reflected relatively superficial relatedness, leaving limited evidence about deeper linguistic history. The paper compares LID models trained across 126 to 4,017 languages and analyzes their representations. Phylogenetic recovery improves at 4K, where the model also captures contact patterns including a Papuan–Oceanic–Australian cluster.

  • Problem

    Prior investigations found that S3Ms encoded relatively superficial language relatedness and recovered limited information about deep language-group history.

    Source: §1 · p. 1

  • Method

    The authors compare four LID models covering 126 to 4,017 languages and analyze language embeddings across 49 languages using hierarchical clustering.

    Source: §1 · p. 1

  • Results

    At 4K, phylogenetic recovery rises from ARI 0.47 to 0.74 and NMI 0.87 to 0.95, with Oceanic, Papuan, and Australian languages forming a macro-cluster.

    Source: §1 · p. 1

  • Takeaways & Limitations

    The findings suggest that scaling linguistic coverage can recover groupings aligned with deep phylogenetic and contact relationships.

    Source: §1 · p. 1, §4 · p. 4

Abstract

from arXiv · show

Similarities between language representations derived from Self-Supervised Speech Models (S3Ms) have been observed to primarily reflect geographic proximity or surface typological similarities driven by recent expansion or contact, potentially missing deeper genealogical signals. We investigate how scaling an S3M-based language identification system from 126 to 4,017 languages reshapes this topology, and find a non-linear effect: phylogenetic recovery stays flat up to the 1K scale, but the 4K model undergoes a qualitative shift, resolving both clear lineages and long-term linguistic contact. Most strikingly, a robust Pacific macro-cluster emerges, grouping genealogically unrelated Papuan, Oceanic, and Australian languages, and we trace its driver to a concentrated encoding that captures shared acoustic signatures such as global energy dynamics. These results suggest that massive S3Ms internalize multiple layers of language history, offering a promising perspective for computational phylogenetics and the study of language contact.

1. Introduction

Prior work found that S3M representations tended to reflect recent phylogenetic and contact relationships, leaving limited information about deeper language history. This study scales LID training coverage to 4,017 languages and reports improved recovery of genealogical and long-term contact relationships, including a Pacific macro-cluster.

1. Introduction

  • Earlier studies found that S3Ms captured relatively superficial relatedness and recovered limited information about deep language-group history.The paper frames speech representations as potentially reflecting linguistic expansion, divergence, and convergence.Source: §1 · p. 1
  • Scaling coverage to 4K improves phylogenetic recovery over models up to 1K, with ARI rising from 0.47 to 0.74 and NMI from 0.87 to 0.95.The 4K model also captures known contact relationships, including the Early Chinese cultural sphere, Persian area, and Dravidian area, with high bootstrap confidence.Source: §1 · p. 1
  • The 4K model groups Oceanic languages with genealogically unrelated Papuan and Australian languages, while Sundaic and Philippine languages form a separate Austronesian cluster.The pattern aligns with Linguistic Melanesia convergence and population-genetic evidence of deep Pacific interaction.Source: §1 · p. 1
  • The authors argue that scaling linguistic coverage qualitatively reshapes representation geometry, recovering patterns aligned with deep phylogenetic and contact relationships.The relationships were otherwise entangled at smaller linguistic scales.Source: §1 · p. 1

2. Methods of Analysis

The analysis uses 49 languages from diverse families and extracts standardized language-level embeddings from four LID models. It evaluates hierarchical clustering against genealogical groupings and examines which acoustic features align with the Pacific cluster.

2.1. Data

  • The dataset comprises 49 languages spanning diverse families, with recordings drawn from DoReCo and FLEURS.DoReCo contributes 14 primarily low-resource languages; 35 typologically diverse languages are selected from FLEURS.Source: §2.1 · p. 1

2.2. Embedding Extraction

  • The four models cover 126, 256, 1,024, and 4,017 languages; 45 of 49 evaluation languages share seen/unseen status between the 1K and 4K models.For each language, final-layer hidden states are double-averaged across audio clips into a centroid, then dimensions are standardized to zero mean and unit variance.Source: §2.2 · pp. 1–2

2.3. Hierarchical Clustering and Evaluation

  • Ward-linkage clustering of standardized centroids is evaluated against subfamily-level genealogical groupings using ARI and NMI across K = 2–20.A file-level bootstrap resamples audio files 1,000 times, recomputes centroids and dendrograms, and assesses clade stability; the Figure 2 caption reports a plateau through 1K and a 4K performance leap peaking at K = 18.Source: §2.3 · p. 2, Figure 2 · p. 2

2.4. Dimension-Level Analysis

  • The dimension-level analysis tests POA against non-POA centroids across 1,280 dimensions and correlates significant dimensions with 30 acoustic features.It applies FDR and Bonferroni corrections, uses two corresponding stringency levels for feature frequency, and independently tests raw-audio feature differences with Mann–Whitney U tests and Cohen’s d.Source: §2.4 · p. 2

3. Results

The 4K model makes a qualitative leap in recovering genealogical structure and contact patterns, while revealing a distinct Papuan–Oceanic–Australian grouping. Further analysis finds that this grouping is spatially distinct and associated with concentrated encoding of broad acoustic dynamics.

3.1. Phylogenetic Recovery

  • Phylogenetic recovery remains stable from 126 to 1K, but at K = 18 the 4K model reaches ARI 0.74 and NMI 0.95.The passage describes this peak as aligning with the actual number of subgroups.Source: §3.1 · p. 2

3.2. Linguistic Contact and Areal Patterns

  • The 4K dendrogram recovers documented contact groupings, and 36 of 37 branches above 50% confidence align with established phylogenetic or areal groupings.These include East Asian, Iranian–Turkic, and Dravidian–Indo-Aryan groupings; Māori clusters near English with 100% bootstrap support, likely reflecting colonial contact and bilingualism.Source: §3.2 · pp. 2–3, Figure 3 · p. 3

3.3. Additional Analysis of the POA Cluster

  • 4K reaches Precision 1.00 across tested K, versus the 1K maximum of 0.92, and F1 0.96, versus the 1K peak of 0.92.The 4K model coherently clusters POA languages and isolates them from other families.Source: Figure 4 · p. 3, §3.3 · pp. 3–4

3.2. Linguistic Contact and Areal Patterns

  • The dendrogram separates Austronesian groups: Sundaic and Philippine languages form Group A, while Oceanic languages join Papuan and Australian languages in Group B.Group B has 57% bootstrap support, with Oceanic and Papuan languages clustering together at 74%; the pattern is linked to Melanesian convergence and Pacific interaction evidence. The Australian–Papuan association is consistent with a long-conjectured relationship.Source: §3.2 · p. 3

3.3. Additional Analysis of the POA Cluster

  • The 4K PCA projection shows cleaner family boundaries and a distinct POA cluster; non-POA DoReCo languages remain with their respective families.The passage contrasts this with overlap and less distinct boundaries in the 1K space.Source: Figure 5 · p. 4, §3.3 · p. 4
  • The 4K model uses fewer significant dimensions for POA grouping and emphasizes global amplitude dynamics, including energy dynamic range.There are 169 versus 257 significant dimensions under FDR and 25 versus 36 under Bonferroni correction. Energy dynamic range is the most frequent 4K feature under Bonferroni correction (28.0%) but does not survive in 1K; raw-signal features align with the model’s priorities.Source: §3.3 · p. 4

4. Conclusion

The study concludes that greater linguistic diversity in S3M training data supports recovery of genealogical subgroups and historical contact patterns. The Papuan–Oceanic–Australian cluster offers an acoustic perspective on regional interaction.

  • Increasing linguistic diversity in S3M training data enables recovery of genealogical subgroups and historical contact patterns.Source: §4 · p. 4
  • The Papuan–Oceanic–Australian cluster points toward long-term convergence in Melanesia and provides an acoustic perspective on regional interactions.Source: §4 · p. 4
  • The findings offer a computational approach to uncovering latent linguistic interactions that are difficult to resolve using traditional methods.Source: §4 · p. 4
Loading 2603.07238v2…