Research
Notes, methods, and lab updates on capturing expert work as training material for frontier AI. Long-form when it warrants it, short when it doesn't.
VOICE-H is a human evaluation benchmark for TTS models: 100 quotes selected from 40,000 real interviews, 17 models against the original human recordings, and the qualitative reasons behind every preference. Updated with 15,217 English comparisons from 901 rater sessions, three new models, and French and Spanish results.
VOICE-H expanded.→Five new TTS models join VOICE-H, taking it to fifteen. Cartesia's Sonic 3.6 takes the top place, Speechify's Simba 3.2 arrives third at a fraction of the price, and the same model can lead in one language and sit mid-table in another.
No posts in this filter yet.