ZTTS2
·
Eval
Reliability and intelligibility of zero-shot text-to-speech, measured across languages
models
—
rows
—
channel
—
view
board
indices
grid
reliability
speaking rate
prosody overlap
task
lang
metric
filter
colour range
clip
min–max
min
max
auto
Reliability — the tail is the whole story
p99 · CI
survival
small multiples
slopegraph
Speaking rate — pacing against natural speech
d · CI
distribution
fastest first
closest first
Prosody realism — distance from ground truth
Capability index
capability
prosody
ranking
over time
parameters