Benchmark › Speech models › Reliability
Reported, not part of the overall score · updated 2026-11-22
Reliability
How often does a request come back with usable audio on the first try?
Word accuracyNumbers, dates and symbolsLong textRepeated wordsOther languagesEmotion controlReliabilitySpeed
Ranking
Share of requests that returned usable audio on the first attempt. Higher is better.
0255075100
Failures in detail
| Model | Requests | Failed twice | Worked on retry | Error returned |
|---|---|---|---|---|
| MiniMax Speech 2.8 Turbo | 257 | 0 | 5 | |
| Gemini 2.5 Pro TTS | 257 | 27 | 7 | no audio in response (OTHER) |
What we test and how it is scored
Test. Every request in the benchmark except the long-text ladder, about 255 per model.
Scoring. A request is clean when the first attempt returns audio that is not silent. Rate-limit responses caused by the test itself are retried and not counted. Failures are retried once; an item that fails twice counts as fully wrong in its own category.
Full item list and rules: methodology.
Think a score is wrong? Check it.
Every test item, scoring rule and known limit is published, and the audio behind each number is on the page.