findsylls: A Language-Agnostic Toolkit for Syllable-Level Speech Tokenization and Embedding

A new open-source toolkit called findsylls aims to standardize syllable-level speech tokenization by unifying classical and neural syllabification methods under a single, language-agnostic interface, according to a paper submitted to arXiv on 27 Mar 2026 [1]. The toolkit, introduced by Héctor Javier Vázquez Martínez, addresses what the paper describes as a fragmented research landscape where syllabification studies rely on "disparate implementations, datasets, and evaluation protocols" [1][2]. findsylls organizes syllable processing into three interoperable modules: envelope computation, frame-level feature extraction, and segmentation algorithms [3][4]. It implements and standardizes widely used methods, including the Sylber model and VG-HuBERT, allowing their components to be recombined for controlled comparisons of representations, algorithms, and token rates [1][4]. Sylber, developed by the Berkeley Speech Group, was previously noted for yielding "extremely short tokens from raw audio (on average, 4.27 tokens/sec) through dynamic tokenization at the syllable granularity" [6]. findsylls builds on this lineage by offering a shared evaluation framework that quantifies nuclei, boundary, and span F1 scores against annotated TextGrids across multiple corpora and languages [4]. The toolkit was demonstrated on English and Spanish corpora and on new hand-annotated data from Kono, an underdocumented Central Mande language [1][2]. The paper reports that benchmarking ten classical and representation-driven configurations across seven corpora revealed "a clear speed–accuracy trade-off: simple envelope baselines can deliver strong nuclei accuracy at high RTFx, while improved boundaries and spans benefit from stronger learned representations and segmentation choices" [4]. Beyond boundary detection, findsylls includes a syllable-level embedding pipeline that combines any segmenter with any feature extractor and aggregates features within each syllable using mean, max, median, or onset–nucleus–coda pooling, producing syllabic tokens suitable for downstream modeling [3][4]. The authors frame syllabic tokenization as "a design space rather than a single fixed method" and position findsylls as shared infrastructure for reproducible evaluation and controlled method composition across languages and domains [4]. A separate recent study proposed ZeroSyl, a training-free method for extracting syllable boundaries directly from a frozen WavLM model, which outperformed prior syllabic tokenizers on syntactic and narrative benchmarks [5]. That work noted that while finer-grained units remain beneficial for lexical tasks, syllabic units exhibit better scaling behavior for syntactic modeling [5]. The findsylls toolkit provides a standardized environment in which such claims can be systematically tested and compared.