A Calculus-Based Framework for Determining Vocabulary Size in End-to-End ASR

2026-05-14 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

A new calculus-based framework has been developed to formally estimate the optimal vocabulary size for end-to-end Automatic Speech Recognition (ASR) systems. Unlike hybrid ASR systems where vocabulary size is clearly defined by linguistic units, end-to-end systems derive tokens from training text corpora, making vocabulary size a critical hyper-parameter. Existing tokenization algorithms like BPE, WordPiece, and ULM require vocabulary size as an input, yet its determination often lacks formal justification in literature and toolkits like ESPNet. This framework builds on prior work by employing curve fitting on training data and applying first and second derivative tests from calculus. The approach was validated using the standard Librispeech corpus, demonstrating that an optimally chosen vocabulary size significantly enhances ASR performance.

Key takeaway

For AI Engineers and Research Scientists optimizing end-to-end ASR systems, applying this calculus-based framework can formally determine the optimal vocabulary size. This approach, validated on the Librispeech corpus, directly improves ASR performance by moving beyond arbitrary hyper-parameter selection. Consider integrating derivative tests into your tokenization pipeline to achieve more robust and efficient model training.

Key insights

A calculus-based framework formally determines optimal vocabulary size for end-to-end ASR systems.

Principles

• Optimal vocabulary size improves ASR performance.
• Calculus derivatives can estimate hyper-parameters.

Method

The method involves curve fitting training data and applying first and second derivative tests from calculus to formally estimate the optimal vocabulary size hyper-parameter for end-to-end ASR.

In practice

• Apply to end-to-end ASR systems.
• Use with tokenization algorithms.

Topics

• End-to-End ASR
• Vocabulary Size Optimization
• Calculus-Based Framework
• Tokenization Algorithms
• Hyper-parameter Tuning

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• A Calculus-Based Framework for Determining Vocabulary Size in End-to-End ASR — A calculus-based framework analytically optimizes ASR vocabulary size, outperforming heuristic methods via differentiable cost functions.
• JSPG: Dynamic Dictionary Filtering via Joint Semantic-Pinyin-Glyph Retrieval for Chinese Contextual ASR — Jointly integrating semantic, pinyin, and glyph features significantly improves Chinese contextual ASR keyword retrieval accuracy.
• Surpassing Scale by Efficiency: A Compact 135M Parameter Foundational LLM Natively Adapted for the Bangla Language — A compact 135M parameter LLM for Bangla achieves 1B-tier performance by efficiently merging tokens, overcoming low-resource script challenges.
• When Multiple Scripts Matter: Evaluating ASR in Clinical Settings — Multiscript variability in non-English clinical ASR requires specialized evaluation and script unification during training for accurate performance.
• Cross-Lingual Empirical Evaluation of Large Language Models for Arabic Medical Tasks — English-centric LLMs exhibit significant performance gaps and reliability issues in Arabic medical tasks.
• Improving low-resource ASR using bilingual fine-tuning with language identification: a cross-linguistic evaluation — Bilingual fine-tuning enhances low-resource ASR, particularly when language identification is accurate or explicitly guided during inference.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.