Peer-reviewed research
Dissecting User-Perceived Latency of On-Device E2E Speech Recognition
Yuan Shangguan et al., Interspeech 2021, 2021 · DOI 10.21437/Interspeech.2021-1887
Read the originalRead on October 19, 2026
- What it is
- A Facebook AI study of how on-device speech recognition models feel to users. Among other measures, it reports the real-time factor of four streaming RNN-T models measured on single-core CPUs in Android devices.
- How much weight it carries
- Peer-reviewed conference paper with open full text from the authors' institution. The speed figures are from one CPU setting on unnamed Android devices, on an in-house voice command dataset, so they are not a benchmark of a named product.
Facts taken from this source
- 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.
“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” (Section 3, Table 1)
- In the paper's Table 1, LSTM RNN-T model B (encoder with 8x320 and 4x640 layers) had a real-time factor of 0.42 when run on a single-core CPU of an Android device, about 2.38 times faster than real time.
“B 8x320+4x640 2x512 57.1 1.47 2.25 26.3 0.42 ... we focus on first-pass, single-core CPU-based model evaluation on Android devices” (Section 3, Table 1)
- In the paper's Table 1, LSTM RNN-T model C (8 encoder layers of 640 cells, 4 predictor layers) had a real-time factor of 0.55 when run on a single-core CPU of an Android device, about 1.82 times faster than real time.
“C 8x640 4x512 91.4 1.57 2.06 36.5 0.55 ... we focus on first-pass, single-core CPU-based model evaluation on Android devices” (Section 3, Table 1)
- In the paper's Table 1, Emformer RNN-T model D had a real-time factor of 0.21 when run on a single-core CPU of an Android device, about 4.76 times faster than real time.
“Emformer-based RNN-T Encoder Cells D 20 x (8 |H|, 512|W |, 5|C|) 2x512 77.4 1.50 2.14 11.8 0.21 ... we focus on first-pass, single-core CPU-based model evaluation on Android devices” (Section 3, Table 1)