[HUGGINGFACE]score: 0.38
SwanBench-Speech Benchmarks Long-Form TTS Consistency and Coherence
May 26, 2026
SwanBench-Speech introduces disentangled evaluation dimensions for long-form speech generation, targeting gaps in existing benchmarks that ignore consistency and coherence across extended contexts and diverse domains like dialog generation.
paper
HOW THIS AFFECTS YOU
●
builderIf you are building or evaluating long-form TTS pipelines, this benchmark offers more granular quality signals than standard metrics like MOS or WER.
●
researcherProvides a more rigorous evaluation framework for long-context TTS models, separating quality dimensions that existing metrics conflate.