Peer-reviewed research
DNN-based speech bandwidth expansion and its application to adding high-frequency missing features for automatic speech recognition of narrowband speech
Kehuang Li et al., Interspeech 2015, International Speech Communication Association (ISCA), 2015
Read the originalRead on October 16, 2026
- What it is
- A peer-reviewed Interspeech paper that compares speech recognition on narrowband (8 kHz) and wideband (16 kHz) audio using the 20,000-word Wall Street Journal task. Wideband gave a word error rate of 8.12% and narrowband 8.67%.
- How much weight it carries
- Peer-reviewed. The comparison is a preliminary experiment on clean read speech with a 2015 recognizer, so it shows the direction and rough size of the gap, not the gap in noisy rooms or on phone calls.
Facts taken from this source
- In the same preliminary Wall Street Journal task, narrowband (8 kHz) speech gave a word error rate of 8.67%.
“wideband speech, with a word error rate (WER) of 8.12%, can always outperform narrowband speech, with a WER of 8.67%” (Introduction, paragraph 2)
- In a preliminary Wall Street Journal recognition task, wideband (16 kHz) speech gave a word error rate of 8.12% with the same recognizer architecture.
“wideband speech, with a word error rate (WER) of 8.12%, can always outperform narrowband speech, with a WER of 8.67%” (Introduction, paragraph 2)