Earnings25: A Comprehensive 500-Hour Speech Benchmark for Finance

Earnings25 is a large-scale benchmark for evaluating automatic speech recognition (ASR) systems on earnings-call audio in the financial domain. The dataset comprises two complementary test sets: • testset-full: 498 hours of full earnings calls from S&P 500 companies (2025 Q4), preserving full conversational structure.• testset-segmented: 46 hours of industry-balanced 5–10 minute segments (290 industry classifications), sampled via stratified sampling from over 2,000 U.S. earnings calls (2025 Q1–Q4). Earnings25 includes aligned transcripts, with timestamps obtained via CTC-based forced alignment, and rich structured metadata including: speaker names and roles (operator, executive, analyst) industry classification company identifiers The benchmark supports fine-grained ASR evaluation including speaker-aware and industry-aware analysis. This dataset accompanies the Interspeech 2026 dataset-track submission:“Earnings25: A Comprehensive 500-Hour Speech Benchmark for Finance”.