Fact · Speech recognition
17.7percent test set error
Graves, Mohamed and Hinton report that deep long short-term memory recurrent networks, trained end-to-end with suitable regularisation, reached a 17.7% test set error on the TIMIT phoneme recognition benchmark.
Checked against its source on October 17, 2026
In the source's own words
“When trained end-to-end with suitable regularisation, we find that deep Long Short-term Memory RNNs achieve a test set error of 17.7% on the TIMIT phoneme recognition benchmark”
Speech Recognition with Deep Recurrent Neural Networks, Alex Graves et al., arXiv (accepted for ICASSP 2013), 2013, Abstract.
Open the original- What it measures
- This was an early result showing that recurrent networks trained end-to-end, with CTC, could compete on speech. The score is for phonemes, the sound units, not for words.
- What it does not show
- TIMIT is a small, clean, phoneme-labelled benchmark from the 1990s. A phoneme error rate cannot be compared with the word error rates of modern systems.