Voice A
Sample A
0:00Loading…
Text-to-speech · blind preference test
Which voice sounds better?
Listen to both clips reading the same prompt. Every active model uses a male voice so comparisons are cleaner. Pick the one you would rather ship in a real product, then tell us why. Model names stay hidden until results are published, and after each vote the next pair is chosen by what would most sharpen the ranking — close races and under-tested voices come up more often.
Prompt
The API uses OAuth, JWT, TLS, and HTTP/2.
← vote
Voice B
Sample B
0:00Loading…
Loudness comparison
Overlaid RMS envelopes on a shared time axis. Differences in length, pace, and emphasis show up here before you hear them.
Sample ASample B
Keyboard shortcuts
←Prefer left (A)
→Prefer right (B)
SpacePlay / pause active
TToggle active player
RRestart active
SToggle spectrogram
CToggle A/B compare
Provisional preference results
Blind Elo leaderboard
Names stay hidden during voting. Results below are preference ratings only, not the hard-text benchmark. This run uses the male-voice pool only. The interval is a provisional 95% uncertainty band from comparison volume, so low-vote rows should be treated as unstable.
Comparisons
0
Start Elo
1500
| Rank | Model | Vendor | Elo | 95% CI | Votes | W-L | Signal |
|---|---|---|---|---|---|---|---|
| Loading ratings... | |||||||
Beyond this leaderboard
These votes train a model that scores any TTS system
The research note turns this preference signal into five comparable axes — preference, Whisper intelligibility, UTMOS naturalness, speaker similarity, and prosody — and asks which objective metric best predicts what you pick here.