fastabx: A library for efficient computation of ABX discriminability

2025-05-05

AI Summary

ABX discriminative evaluation—widely adopted for assessing phoneme discriminability of speech self-supervised representations—is hindered by tooling scarcity and computational inefficiency. To address this, we introduce FastABX: the first general-purpose, modular, and highly optimized ABX framework, implemented in Python with Cython acceleration, memory-mapped I/O, and batched cosine/Euclidean distance computation. It supports flexible configuration and arbitrary representation inputs. FastABX accelerates ABX task construction and distance computation by over an order of magnitude versus baseline implementations, enabling real-time processing of datasets with millions of samples. Its generic architecture transcends speech-specific constraints, establishing a scalable infrastructure for discriminative evaluation of multimodal representations. The open-source implementation has been integrated into multiple representation learning projects, serving as a community standard for rigorous, efficient ABX benchmarking.

Technology Category

• Machine Learning: Multimodal Learning
• Natural Language Processing: Speech
• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

We introduce fastabx, a high-performance Python library for building ABX discrimination tasks. ABX is a measure of the separation between generic categories of interest. It has been used extensively to evaluate phonetic discriminability in self-supervised speech representations. However, its broader adoption has been limited by the absence of adequate tools. fastabx addresses this gap by providing a framework capable of constructing any type of ABX task while delivering the efficiency necessary for rapid development cycles, both in task creation and in calculating distances between representations. We believe that fastabx will serve as a valuable resource for the broader representation learning community, enabling researchers to systematically investigate what information can be directly extracted from learned representations across several domains beyond speech processing. The source code is available at https://github.com/bootphon/fastabx.

Problem

• Lack of efficient tools for ABX discriminability computation
• Limited broader adoption of ABX measure in representation learning
• Need for rapid ABX task creation and distance calculation

Innovation

• High-performance Python library for ABX tasks
• Efficient computation of ABX discriminability
• Framework for constructing any ABX task