Deep Speech: Scaling up End-to-End Speech Recognition
2014 · paper · AI and machine learning
Describes Deep Speech, a speech recognition system trained end-to-end with a large recurrent network, multiple GPUs and synthesized training data, without hand-engineered components. It reports 16.0% error on Switchboard Hub5'00 and better performance than commercial systems in noise.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
2 findings
Reference or label The text says "The model from Vesely et al. (DNN-GMM sMBR) [44]", but the Table 3 row it refers to is "Vesely et al. (DNN-HMM sMBR) [44]".
- p. 7 Section 5.1, last paragraph
The model from Vesely et al. (DNN-GMM sMBR) [44]
- p. 8 Table 3, row 2
Vesely et al. (DNN-HMM sMBR) [44]
Judge's note: The p7 crop has 'DNN-GMM sMBR', and the same sentence describes 'a typical hybrid DNN-HMM system'. The p8 Table 3 has 'DNN-HMM sMBR' with Full 18.4, which is the baseline behind the stated 2.4% absolute gain. Table 3 has no DNN-GMM row.
Second check: another model agreed
Report a problem with this finding
Reference or label Two different papers in the reference list carry the same identifier "abs/1406.7806, 2014. http://arxiv.org/abs/1406.7806": [19] (Hinton et al., co-adaptation/dropout) and [28] (Maas et al., DNN acoustic model size).
- p. 11 Reference [19]
ing neural networks by preventing co-adaptation of feature detectors. abs/1406.7806, 2014.
- p. 11 Reference [28]
abs/1406.7806, 2014. http://arxiv.org/abs/1406.7806.
- p. 11 Reference [19]
Improv-ing neural networks by preventing co-adaptation of feature detectors. abs/1406.7806, 2014.
- p. 11 Reference [19]
abs/1406.7806, 2014.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that 1406.7806 is the Maas et al. paper and that the dropout preprint is arXiv 1207.0580 (2012). The page 11 image shows the same ID and URL printed in both [19] and [28]. Two distinct papers cannot share one arXiv ID, so at least one entry is wrong; the authors need to say which.
Second check: another model agreed
Report a problem with this finding
6 minor (typos and the like; most checked by the judge only)
Wording "Here W_k^(6) and b_k^(6) denote the k’th column of the weight matrix". In exp(W_k^(6) h_t^(5) + b_k^(6)), with layers written as W h, W_k^(6) has to be the k'th row, or a transpose is needed.
- p. 3 Text below the softmax equation
denote the k’th column of the weight matrix and k’th bias, respectively.
Judge's note: The p3 crop shows the softmax written as W_k^(6) h_t^(5). Layer 5 is written g(W^(5) h_t^(4) + b^(5)), which is a matrix times a column vector. Under that convention the score for output k is the k'th row of W^(6) times h. No transpose is shown anywhere. This is a notation slip; the model itself is unaffected.
Second check: another model agreed
Report a problem with this finding
Wording "first introduced to speech pipelines in the early 1990s [1, 34, 11]" cites [11], which the reference list dates "IEEE, 1999".
- p. 9 Section 6, first sentence
first introduced to speech pipelines in the early 1990s [1, 34, 11]
- p. 10 Reference [11]
In ICASSP, volume 2, pages 1013–1016. IEEE, 1999.
Judge's note: The p10 crop shows [11] as Ellis and Morgan, 'Size matters: An empirical study...', ICASSP 1999. The p11 page shows [34] dated 1994. The 1999 study can support the general point but cannot be one of the early-1990s works that first introduced these models. This is loose attribution, not a factual error.
Second check: another model agreed
Report a problem with this finding
Typo Stray period in "Lee et. al. [27]"; it should read "et al."
- p. 2 Introduction, paragraph 2
Lee et. al. [27]
Judge's note: The p2 crop shows 'Lee et. al. [27]'. The same paragraph writes 'Coates et al. [7]'.
Second check: none (judge only)
Report a problem with this finding
Typo Stray article in "a similar solution was used by the Sutskever et al. [42]"; 'the' should be deleted.
- p. 5 Section 3.1, last sentence
used by the Sutskever et al. [42]
- p. 5 Section 3.1 last sentence
was used by the Sutskever et al. [42]
Judge's note: The p5 crop shows the wording as quoted, at the end of Section 3.1.
Second check: none (judge only)
Report a problem with this finding
Typo Missing word in "our system is able perform a full pass over the 2300 hours of data"; it should read "able to perform".
- p. 7 Section 5.1, paragraph 2
our system is able perform a full pass
- p. 7 Section 5.1, paragraph 2
our system is able perform a full pass over the 2300 hours of data
Judge's note: The p7 crop shows the wording as quoted, so this is not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo "trained on only 300 hour switchboard" uses lower case, while the paper capitalises 'Switchboard' everywhere else.
- p. 7 Section 5.1
trained on only 300 hour switchboard
Judge's note: The p7 crop (C005-a) shows 'only 300 hour switchboard'. A text search finds every other use capitalised, including 'trained on 300 hour Switchboard' on the same page.
Second check: none (judge only)