Fact · Speech recognition
250 to 300times faster than real time
Smallest AI reports that Pulse Pro transcribes a single two-hour audio file at 250 to 300 times real time on one NVIDIA L40S GPU when word timestamps are off, so about 275 times on average (midpoint).
Quote found word for word on the live source, last checked October 19, 2026
In the source's own words
“Long-form audio on L40S (single 2 hr file) Mode Throughput (RTFx) 2 hr file latency No word timestamps 250–300× ~24–29 sec”
Parallelism and Latency (Smallest AI self-hosted Pulse STT), Smallest AI, n.d., Parallelism and Latency, Pulse Pro, Long-form audio on L40S.
Open the original- What it measures
- The page states that the figures come from internal benchmarks on a single GPU host, measured on an L40S (48 GB) for reference. The recommended production GPU is an L4, which the page says is slower. The page gives the 2-hour file latency as about 24 to 29 seconds for this row.
- What it does not show
- A vendor's internal benchmark, single file, English not stated, steady-state on one GPU. The page warns that throughput depends on hardware revision, driver and audio, and that the L4 it recommends is slower. The midpoint 275 is calculated from the range.
In the data
A row of Speech Recognition Speed: Real-Time Factor Data, as Smallest AI, 2 hr file, no timestamps: 250 to 300 times faster than real time.
- System
- Smallest AI Pulse Pro (self-hosted)
- Kind
- Cloud service
- Hardware
- NVIDIA L40S GPU (48 GB)
- Test set
- Single 2-hour file, no word timestamps
- Note
- Midpoint of the published range 250 to 300 times real time.