LenguASR / Documentation
Performance notes
Pentium Gold G6600 · baseline runtime · Q8 Nemotron + Silero VAD
These tests used the baseline runtime (no AVX2), the Q8 Nemotron model and Silero VAD with the 11-second JFK reference recording. Each setting was run three times after warm-up, with one VAD thread. The recording was fed in faster than real time, followed by silence, then finalized.
The results show how quickly the recognizer processes audio on its own. They don’t measure gameplay frame time or the delay between speaking into the mic and getting a final transcript.
| ASR threads | Context | Median | Range | RTF |
|---|---|---|---|---|
| 1 | Model / 320 ms | 13.483 s | 13.416-13.663 s | 1.226 |
| 2 | Model / 320 ms | 8.474 s | 8.461-8.798 s | 0.770 |
| 3 | Model / 320 ms | 7.308 s | 7.259-7.642 s | 0.664 |
| 4 | Model / 320 ms | 8.378 s | 8.093-8.722 s | 0.762 |
| 1 | Low latency / 160 ms | 14.547 s | 14.530-15.588 s | 1.322 |
| 2 | Low latency / 160 ms | 9.919 s | 9.560-10.403 s | 0.902 |
| 3 | Low latency / 160 ms | 7.925 s | 7.900-8.265 s | 0.720 |
| 4 | Low latency / 160 ms | 9.470 s | 9.324-10.019 s | 0.861 |
RTF is the processing time divided by the length of the recording. A value below 1 means the recognizer processed the recording in less time than it takes to play it.
Memory and live response time
The tests measured about 779 MB resident memory, 806 MB private commit and 934 MB peak commit. Allow 1.25 GB of additional available memory for each active recognizer.
Support
Need help?
Send us your engine version, CPU, microphone, listening mode, relevant settings and LenguASR logs.