Comprehensive Benchmarking of Long-Form Speech Generation in Diverse Scenarios

Researchers introduced Swanbench-Speech, a benchmark for long-form speech and dialog generation across 17 realistic scenarios.

The dataset contains 1,101 samples and uses seven automated metrics to evaluate acoustics, semantics, and expressiveness.

Results show current speech models still struggle with expressive delivery, consistency, and maintaining hierarchical structure.

The benchmark is designed to support more reliable, broader evaluation of long-form speech quality.