Peer-reviewed research
Spike-Triggered Non-Autoregressive Transformer for End-to-End Speech Recognition
Zhengkun Tian et al., Interspeech 2020, 2020 · DOI 10.21437/Interspeech.2020-2086
Read the originalRead on October 19, 2026
- What it is
- A paper from the Chinese Academy of Sciences that proposes a non-autoregressive transformer for Mandarin speech recognition, trained and tested on AISHELL-1. Its experiments report the real-time factor of the model at several CTC trigger thresholds, measured on a single GeForce GTX TITAN X GPU.
- How much weight it carries
- Peer-reviewed conference paper with open full text. The speed figures come from one GPU, one dataset and one decoding setup, with two significant digits (0.0056), so the converted speeds are only approximate. Speed was measured with utterances decoded one at a time.
Facts taken from this source
- In a test on AISHELL-1 run on a GeForce GTX TITAN X GPU, the spike-triggered non-autoregressive model at trigger threshold 0.1 had a real-time factor of 0.0059, about 169 times faster than real time.
“on a GeForce GTX TITAN X 12G GPU. ... Threshold β 0.1 0.3 0.5 0.7 Performance 7.88 7.67 7.77 7.70 Seconds 212.04 202.59 200.62 198.44 RTF 0.0059 0.0056 0.0055 0.0054” (Section 3.3, Table 2)
- In a test on AISHELL-1 run on a GeForce GTX TITAN X GPU, the spike-triggered non-autoregressive model at trigger threshold 0.3 had a real-time factor of 0.0056, about 179 times faster than real time.
“on a GeForce GTX TITAN X 12G GPU. ... Threshold β 0.1 0.3 0.5 0.7 Performance 7.88 7.67 7.77 7.70 Seconds 212.04 202.59 200.62 198.44 RTF 0.0059 0.0056 0.0055 0.0054” (Section 3.3, Table 2)
- In a test on AISHELL-1 run on a GeForce GTX TITAN X GPU, the spike-triggered non-autoregressive model at trigger threshold 0.5 had a real-time factor of 0.0055, about 182 times faster than real time.
“on a GeForce GTX TITAN X 12G GPU. ... Threshold β 0.1 0.3 0.5 0.7 Performance 7.88 7.67 7.77 7.70 Seconds 212.04 202.59 200.62 198.44 RTF 0.0059 0.0056 0.0055 0.0054” (Section 3.3, Table 2)
- In a test on AISHELL-1 run on a GeForce GTX TITAN X GPU, the spike-triggered non-autoregressive model at trigger threshold 0.7 had a real-time factor of 0.0054, about 185 times faster than real time.
“on a GeForce GTX TITAN X 12G GPU. ... Threshold β 0.1 0.3 0.5 0.7 Performance 7.88 7.67 7.77 7.70 Seconds 212.04 202.59 200.62 198.44 RTF 0.0059 0.0056 0.0055 0.0054” (Section 3.3, Table 2)