Fact · Speech recognition
0.53real-time factor
In the paper's Table 1, LSTM RNN-T model A (8 encoder layers of 640 cells) had a real-time factor of 0.53 when run on a single-core CPU of an Android device, about 1.89 times faster than real time.
Checked against its source on October 19, 2026. On October 19, 2026 the quote was no longer on the live page, so the source may have changed since.
In the source's own words
“we focus on first-pass, single-core CPU-based model evaluation on Android devices ... P50 WER FLOPS RTF layers layers (MB) UPL(s) mil P50 A 8x640 2x512 69.9 1.60 1.94 33.8 0.53”
Dissecting User-Perceived Latency of On-Device E2E Speech Recognition, Yuan Shangguan et al., Interspeech 2021, 2021, Section 3, Table 1.
Open the original (DOI 10.21437/Interspeech.2021-1887)- What it measures
- Table 1 compares four streaming RNN-T models by real-time factor, the time needed to process audio frames divided by the length of the audio. The paper says it evaluated single-core CPU inference on Android devices.
- What it does not show
- The Android device model is not named, and the speed is for one CPU core on phones, not for a desktop or server. The test data is an in-house voice command dataset.
In the data
A row of Speech Recognition Speed: Real-Time Factor Data, as Shangguan et al. 2021, model A: 2 times faster than real time.
- System
- LSTM RNN-T model A
- Kind
- Research system
- Hardware
- Single-core CPU, Android device
- Test set
- In-house voice command dataset
- Year
- 2021
- Note
- 1 divided by the real-time factor 0.53 is 1.89 times faster than real time.