Blind Voice Arena
Same prompt, two hidden providers, multiple raters, A/B/tie. VoicePilot avoids pretending a single editor's 4.2/5 rating is objective.
Audio samples
—
Reviewable
—
Open pairs
—
Completed pairs
—
Votes
—
Why this is stronger
Pairwise listening reduces provider-brand bias and makes disagreements visible. Tie is valid, and no winner appears until minimum-vote thresholds are reached.
Four evidence layers
- Generation integrity
- Objective checks
- Blind listening
- Revision cost and time
Current state
Loading arena backend…
Sarvam's first public Bulbul V3 run remains generation evidence only because no reviewable audio artifact was captured. That is intentional.