Inference Speed Details

Run 20261029-070100; 10 inference steps; concurrency 1; warmed requests. RTF is wall-clock generation time divided by generated audio duration. Lower is better. Each call uses a stable seed derived from its call ID so compatible runtimes can be compared exactly.

nano

Mode Calls Median RTF p95 RTF Speed
design 240 0.271 0.320 3.71x
clone 72 0.242 0.306 4.05x

Languages

Language Calls Median RTF p95 RTF Speed
Arabic 13 0.267 0.305 3.98x
Chinese 13 0.273 0.304 3.62x
Danish 13 0.272 0.336 3.81x
Dutch 13 0.267 0.304 3.74x
English 13 0.269 0.314 3.81x
Finnish 13 0.307 0.400 3.20x
French 13 0.264 0.319 3.79x
German 13 0.256 0.307 3.95x
Greek 13 0.286 0.338 3.37x
Hindi 13 0.293 0.341 3.70x
Indonesian 13 0.248 0.264 4.04x
Italian 13 0.280 0.302 3.80x
Japanese 13 0.224 0.283 4.26x
Korean 13 0.265 0.294 3.78x
Malay 13 0.259 0.319 3.84x
Polish 13 0.254 0.297 3.94x
Portuguese 13 0.254 0.287 3.95x
Russian 13 0.242 0.295 4.06x
Spanish 13 0.266 0.307 3.90x
Swedish 13 0.232 0.268 4.24x
Tagalog 13 0.264 0.305 3.97x
Thai 13 0.254 0.281 4.10x
Turkish 13 0.257 0.281 4.09x
Vietnamese 13 0.274 0.294 3.83x