Top AI Product

SwanBench-Speech is a benchmark for the part of speech generation that single-utterance demos hide: long form. Most TTS evaluations score short clips on naturalness and prosody. What breaks when you push to minutes of continuous speech — drift, identity loss, expressive monotony — they tend not to measure.

## 1,101 samples across 17 scenarios

The benchmark covers 1,101 samples spanning 17 common speech scenarios, evaluating acoustic quality, semantic faithfulness, and expressiveness — the three places long-form synthesis tends to crack. Seventeen scenarios spread the test surface beyond audiobook reads to dialogue, news, lectures, narration, and the kinds of situational shifts that catch a model leaning too heavily on whatever style it picked first.

## Why it matters

Long-form synthesis is the production bar for podcasts, audiobooks, dubbed video, character voices in games, and any voice agent that has to maintain identity across a real conversation. A clip that sounds great at five seconds can fall apart at five minutes — pitch wanders, energy flattens, the speaker stops sounding like the speaker. Benchmarks shape what labs train against. A focused long-form, multi-scenario suite gives the field something specific to push on, instead of optimising the same MOS scores on short samples and quietly stalling on the use cases that actually pay.