IMS-Speech: A Speech to Text Tool

Ngoc Thang Vu; Pavel Denisov

REVIEW 3 major objections 5 minor 32 references

IMS-Speech: A Speech to Text Tool

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read One generic speech recognizer, trained on many public corpora, matches or beats task-specific systems on English and German benchmarks.

desk verdict Useful ASR tool paper with competitive results on standard benchmarks, but the 'generic' claim outruns the evidence since all test sets are within training domains. read the letter →

arxiv 1908.04743 v1 pith:3GTNNPFZ submitted 2019-08-13 cs.CL cs.SDeess.AS

Pavel Denisov , Ngoc Thang Vu This is my paper · ORCID

classification cs.CLcs.SDeess.AS

keywords speechrecognitionspeech-to-textend-to-endASRgenericacousticmodelEnglishGermanworderrorratewebservice

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper presents IMS-Speech, a web-based speech-to-text service for English and German, and argues that a single "generic" recognition model can serve many different recording conditions. The authors train one English and one German end-to-end model on a large mixture of public speech corpora, including read, spontaneous, meeting, and distant audio, and compare the resulting word error rates (WER) with published task-specific systems. They report that the generic models are competitive on all tested benchmarks and better on several, including Tuda-De and Verbmobil German data. The goal is to let researchers without speech expertise obtain lexical information from audio recordings through a simple web interface.

What carries the argument

The system combines two trained components: a speech-activity detector that cuts an upload into short speech segments, and an end-to-end recognizer that maps each segment's audio features to a transcription. The recognizer is an attention-based encoder-decoder network trained with a hybrid objective that adds Connectionist Temporal Classification (CTC) loss to the usual attention loss, so the model learns the monotonic alignment between audio and text. Outputs are sub-word units, character-level pieces that can be combined to form any word, which keeps the vocabulary compact and avoids out-of-vocabulary failures. The generic behavior comes from the training mixture: about 2,277 hours of English and 3,797 hours of German covering read, spontaneous, mixed, and distant speech, with the German data augmented to ten times its original size.

What would settle it

Take audio recorded in conditions absent from the training corpora, such as telephone codec speech, overlapping multi-speaker conversations, or background music, and transcribe it with the public IMS-Speech service. If word error rates on those clips are substantially higher than the clean read and spontaneous speech benchmarks reported in Table 2, the claim of generic competitive accuracy across diverse conditions is falsified for exactly those conditions.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that pooling heterogeneous public speech data into one model does not cost accuracy: the generic systems achieve WER of 3.8% on WSJ eval'92, 4.4% on LibriSpeech test-clean, and 17.4% on AMI IHM, alongside German results of 11.1% on Tuda-De dev and 6.7% on Verbmobil 1 dev. Against the best published specialised systems cited in the paper, the generic model is slightly worse on some English benchmarks but clearly better on the German ones. The authors attribute the occasional advantage to better generalization from a larger amount of training data, and they take the overall pattern as evidence that a single generic service can provide transcriptions of competitive accuracy across tasks and conditions.

Load-bearing premise

The claim that the models are generic is tested only on corpora that also supplied training data, so accuracy on genuinely unseen recording conditions is not demonstrated; if users upload audio unlike those corpora, the reported word error rates may not hold.

Editorial extensions

If this is right

A non-specialist researcher can upload English or German audio and receive a usable transcript without building, training, or tuning an ASR pipeline.
On the German benchmarks Tuda-De and Verbmobil, the generic model beats the published task-specific results, suggesting that scale and diversity of training data can compensate for lack of task-specific adaptation.
The beam-width experiments show that reducing search width from 20 to 5 raises WER only slightly, from 12.0% to 13.7% on Tuda-De, while cutting computation roughly in half, so users can trade a little accuracy for speed.
Batched GPU inference processes one second of audio in about 0.7 seconds at the full beam width, making web-based use practical.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the reported test sets all come from corpora used in training, the strongest unmeasured implication is whether the model generalizes to genuinely new recording conditions; a held-out evaluation with far-field, telephone, or overlapping speech would settle that.
If the generic claim holds, a practical downstream consequence is that large archives of oral history, interviews, and lecture recordings in English and German could be transcribed at scale without per-collection model training.
The comparison with a commercial speech API suggests open generic models are approaching commercial services on clean read and spontaneous speech; the same comparison on noisy or distant audio would show where the boundary lies.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Here's my take on 1908.04743. It's a decent system paper: the authors built a web-based German/English transcription service by gluing together Kaldi SAD, ESPnet end-to-end ASR, and public corpora, and they report competitive WER numbers on a range of standard benchmarks. The main contribution is the trained models and the working tool, not new algorithms. For non-technical researchers who need transcripts of, say, oral history interviews or lecture recordings, this kind of drop-in service is genuinely useful, and the paper gives enough detail on architecture, training data, and hyperparameters to reproduce or adapt the models.

What the paper does well: it uses a sensible mix of existing open-source components, trains on a large and varied set of corpora, and evaluates on several benchmarks across both languages. The WER numbers look competitive—often better than the published SOTA for the German datasets they cite, and in the same ballpark for English. They also provide a useful speed/WER tradeoff analysis for beam width and batching. The writing is clear and straightforward.

The soft spots are real but not fatal. The biggest one: every evaluation set in Table 2 is drawn from a corpus that also appears in the training list in Table 1. So the reported WERs demonstrate good within-domain generalization across multiple domains, but they do not show how the system performs on genuinely unseen recording conditions—which is what 'generic' implies for a tool that accepts arbitrary user uploads. The sentence in Section 5.1 about 'possibly due to better generalization from larger amount of training data' is an untested guess. Without an evaluation on a domain absent from training, or at least a sharper statement like 'competitive on the training domains,' the central claim overreaches. Also, the Google API comparison excludes a large chunk of Verbmobil (only 53% scored), which makes that particular comparison less convincing. And there are no confidence intervals, and the SOTA numbers come from heterogeneous sources, so treat the precise WER comparisons as indicative rather than definitive.

Overall: this is a solid, useful tool paper with an overbroad claim. It deserves peer review—the tool is real, the models are trained on public data, and the evaluation is extensive enough to be discussed and improved. I'd recommend sending it to a relevant venue (e.g., Interspeech or LREC). The authors should be asked to either provide an out-of-domain eval or soften the claim.

Referee Report

3 major / 5 minor

Summary. The paper presents IMS-Speech, a web-based speech-to-text tool for English and German, built on Kaldi for speech activity detection and feature extraction and on ESPnet for end-to-end ASR with CTC/attention joint training and subword outputs. The English acoustic model is trained on about 2,277 hours and the German model on about 3,797 hours of public speech corpora, with external language models trained on the corresponding training transcriptions and additional German text. The paper reports word error rates on standard benchmarks (WSJ, LibriSpeech, TED-LIUM, AMI, Tuda-De, Verbmobil), compares against published state-of-the-art numbers and against Google's speech API on a subset of utterances, and shows batched GPU inference speedups in Table 3. The central claim is that these generically trained models achieve competitive accuracy across a diverse set of tasks and conditions.

Significance. If the claims are supported, the tool would be a useful free academic resource for non-specialist researchers who need transcriptions of spoken-language materials. The paper's main contribution is engineering: a reproducible pipeline combining a pretrained Kaldi SAD model, ESPnet-based end-to-end acoustic modeling, and external subword language models over a large multilingual training corpus, wrapped in a web interface. The reported results show that a single generic system can reach competitive WERs on several benchmark test sets, and Table 3 usefully documents a substantial real-time-factor improvement from GPU batching at equal WER. These strengths are tempered by the fact that every evaluation set shares its recording conditions with the training data, so the 'generic, cross-condition' generalization claim is not directly evidenced.

major comments (3)

[§5.1, Tables 1 and 2] The central claim that the systems are generic and achieve 'competitive accuracy on a diverse set of tasks and conditions' is not fully supported by the reported evaluations, because every evaluation set in Table 2 comes from a corpus that also appears in the training list in Table 1: WSJ eval'92, LibriSpeech test-clean/test-other, TED-LIUM 3 test, AMI IHM/SDM/MDM eval, Tuda-De dev/test, and Verbmobil 1 dev/test. The external language model is additionally trained on 'transcriptions from the training speech databases' (Section 4), so the textual and acoustic conditions of the test sets have been seen during training. Held-out speakers within the same corpora demonstrate within-domain speaker generalization, not the ability to handle genuinely new recording conditions. The conclusion's explanation that gains come 'possibly due to better generalization from larger amount of training data' is likewise an untested hypothesis. The authors should either narrow the claim to 'competitive on the domains represented in the training data' or provide an out-of-domain evaluation on a corpus whose recording conditions, microphones, and speaking styles are absent from the training set.
[§5.2, Tables 4 and 5] The Google API comparison is not a clean apples-to-apples comparison because the evaluation excludes all utterances for which Google produced digit-containing transcriptions or empty transcriptions. The reported coverage is 93% (LibriSpeech test-clean), 92% (test-other), 94% (Common Voice), 85% (Tuda-De), and only 53% (Verbmobil 1 test), so a large fraction of the hardest or most domain-specific utterances is removed. Since numeric transcriptions are a natural source of mismatch between systems, selectively filtering them can bias the WER gap in favor of IMS-Speech. The finding that Google has higher WER on the remaining utterances may still be informative, but the text should report the proportion of excluded utterances for each system separately, present results with standard numeric normalization, or include the full set to make the comparison interpretable.
[§5.1, Table 2] The 'state of the art' column mixes results from heterogeneous sources that were not necessarily produced under the same evaluation conditions, decoding setups, or training-data restrictions. For example, the TED-LIUM 3 test reference (6.7% WER) and the CAPIO LibriSpeech references come from systems with different architectures and possibly different external resources, while the AMI numbers point to Kaldi recipe logs. Because the manuscript uses the comparison to support the 'compete with task-specific systems' claim, the text should at least state, for each reference, whether the compared system uses task-specific adaptation, additional training data, or different scoring, or should accompany the comparison with confidence intervals or multiple evaluation runs so that small WER differences (e.g., WSJ 3.8 vs. 3.5) are not overinterpreted.

minor comments (5)

[§5.2] The section heading contains a typo: 'Comparisions' should be 'Comparisons'.
[§4] In the sentence 'External LM for the German system is trained on all transcriptions form the training speech databases', 'form' should be 'from'.
[§5.2] The sentence 'The numbers suggest that that Google API models may be optimized...' contains a doubled 'that'; it should read 'that Google API models'.
[§3, Table 3] Table 3 reports WER and real-time factor for beam widths 5, 10, 15, and 20, but the text does not state whether these runs use the same batch size on GPU and CPU or whether WER differences across beam widths are statistically meaningful; a sentence clarifying the evaluation setup would help.
[References] Reference [13] appears as 'V ery deep convolutional networks...' due to a typographical artifact; please correct the spacing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper is a system description with direct empirical WER measurements on standard benchmarks.

full rationale

The paper reports an ASR tool and evaluates its performance by measuring word error rate on established test sets (WSJ eval'92, LibriSpeech test-clean/test-other, TED-LIUM 3 test, AMI eval splits, Tuda-De dev/test, Verbmobil 1 dev/test). There is no derivation chain in which a predicted quantity is constructed from its own definition, and no fitted parameter is renamed as a prediction. Training details (network architecture, CTC/attention loss, LM scaling factors, subword vocabulary size) are standard model-fitting choices tuned on validation data, which is normal empirical practice rather than circular reasoning. The claim that the systems are 'generic' and 'competitive' is a generalization from measured results, not a quantity derived from the model's own equations. The concern that all test sets come from corpora used in training is a legitimate limitation about external validity and generalization to unseen conditions, but it is a correctness or evaluation-design issue, not circularity within the paper's logic. Citations to Kaldi, ESPnet, and other toolkits are external software components, not self-citations carrying unverified premises. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends on standard ASR evaluation assumptions and a few hyperparameters tuned on validation data. No new theoretical entities or ad hoc constructs are introduced.

free parameters (6)

Sub-word vocabulary size = 100
Evaluated sizes between 50 and 5000; 100 gave better results for both languages (Section 3). This is a hyperparameter chosen based on validation performance.
CTC weight lambda = 0.5
Set to 0.5 for both training and decoding (Section 3). Chosen by the authors, affects the balance of CTC and attention losses.
LM scaling factor gamma (English) = 0.5
Set during decoding for the English system (Section 3). Tuned on validation data.
LM scaling factor gamma (German) = 1.1
Set during decoding for the German system (Section 3). Tuned on validation data.
Beam size = 20
Used for main results; Table 3 shows the trade-off between beam size and speed and WER (Section 5.1).
Maximum speech segment duration = 30 seconds
Extended from the default 10 seconds in the ASPIRE recipe (Section 3).

assumptions (4)

domain assumption Word error rate (WER) is a valid and sufficient metric for transcription accuracy.
The paper uses WER exclusively to compare systems; no human evaluation or confidence intervals are provided (Section 5).
domain assumption The pretrained Kaldi SAD model, trained on Fisher-English, generalizes to the German and English recordings uploaded by users.
The SAD model is downloaded from Kaldi models m4 and used as-is for both languages (Section 3).
domain assumption The test sets used in Table 2 are comparable despite coming from different sources with potentially different scoring protocols.
State-of-the-art WER numbers are collected from various papers and GitHub results without a common scoring setup (Section 5.1).
domain assumption The multi-corpus training data is representative of the diverse conditions the tool targets.
The paper claims generic performance, but the test sets are drawn from the same corpora as training (Sections 4 and 5).

how reviews work

0 comments

Cite this review

Pith. "Pith review of IMS-Speech: A Speech to Text Tool." pith.science (2026). https://pith.science/paper/3GTNNPFZ

@misc{pith2026190804743,
  author       = {Pith},
  title        = {Pith review of: IMS-Speech: A Speech to Text Tool},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GTNNPFZ}},
  note         = {Machine review of arXiv:1908.04743}
}

read the original abstract

We present the IMS-Speech, a web based tool for German and English speech transcription aiming to facilitate research in various disciplines which require accesses to lexical information in spoken language materials. This tool is based on modern open source software stack, advanced speech recognition methods and public data resources and is freely available for academic researchers. The utilized models are built to be generic in order to provide transcriptions of competitive accuracy on a diverse set of tasks and conditions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 22 canonical work pages

[1]

G HOSHAL , G

P OVEY, D., A. G HOSHAL , G. B OULIANNE , L. B URGET , O. G LEMBEK , N. G OEL , M. H ANNEMANN , P. M OTLICEK , Y. Q IAN , P. S CHWARZ ET AL .: The Kaldi speech recognition toolkit. In Proc. of ASRU. 2011

work page 2011
[2]

H ORI , S

W ATANABE , S., T. H ORI , S. K ARITA , T. H AYASHI, J. N ISHITOBA , Y. U NNO , N. E. Y. SOPLIN , J. H EYMANN , M. W IESNER , N. C HEN ET AL .: ESPnet: End-to-End Speech Processing Toolkit. arXiv preprint arXiv:1804.00015, 2018

arXiv 2018
[3]

H ANAZAWA, G

W AIBEL , A., T. H ANAZAWA, G. H INTON , K. S HIKANO , and K. J. L ANG : Phoneme recognition using time-delay neural networks . In Readings in speech recognition . 1990

work page 1990
[4]

M ANOHAR , D

G HAHREMANI , P., V. M ANOHAR , D. P OVEY , and S. K HUDANPUR : Acoustic Modelling from the Signal Domain Using CNNs . In Proc. of Interspeech. 2016

work page 2016
[5]

IEEE Transactions on Information Theory , 1967

V ITERBI , A.: Error bounds for convolutional codes and an asymptotically optimum de- coding algorithm. IEEE Transactions on Information Theory , 1967

work page 1967
[6]

C HO, and Y

B AHDANAU , D., K. C HO, and Y. B ENGIO : Neural machine translation by jointly learn- ing to align and translate . arXiv preprint arXiv:1409.0473, 2014

arXiv 2014
[7]

F ERNÁNDEZ , F

G RAVES, A., S. F ERNÁNDEZ , F. G OMEZ , and J. S CHMIDHUBER : Connectionist tempo- ral classiﬁcation: labelling unsegmented sequence data wi th recurrent neural networks . In Proc. of ICML. 2006

work page 2006
[8]

H ADDOW , and A

S ENNRICH , R., B. H ADDOW , and A. B IRCH : Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015

arXiv 2015

Show all 32 references

[9]

I RIE , R

Z EYER , A., K. I RIE , R. S CHLÜTER , and H. N EY: Improved training of end-to-end at- tention models for speech recognition . arXiv preprint arXiv:1805.03294, 2018

2018 arXiv
[10]

arXiv preprint arXiv:1804.10959, 2018

K UDO , T.: Subword Regularization: Improving Neural Network Transla tion Models with Multiple Subword Candidates. arXiv preprint arXiv:1804.10959, 2018

2018 arXiv
[11]

M ILLER , and K

C IERI , C., D. M ILLER , and K. W ALKER : The Fisher Corpus: a Resource for the Next Generations of Speech-to-Text. In LREC. 2004

2004
[12]

P EDDINTI , D

K O, T., V. P EDDINTI , D. P OVEY , M. L. S ELTZER , and S. K HUDANPUR : A study on data augmentation of reverberant speech for robust speech r ecognition. In Proc. of IEEE ICASSP. 2017

2017
[13]

S IMONYAN , K. and A. Z ISSERMAN : V ery deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

2014 arXiv
[14]

F ERNÁNDEZ , and J

G RAVES, A., S. F ERNÁNDEZ , and J. S CHMIDHUBER : Bidirectional LSTM networks for improved phoneme classiﬁcation and recognition . In International Conference on Artiﬁcial Neural Networks , pp. 799–804. Springer, 2005

2005
[15]

H OCHREITER , S. and J. S CHMIDHUBER : Long short-term memory. Neural computation, 9(8), pp. 1735–1780, 1997

1997
[16]

D.: ADADELTA: an adaptive learning rate method

Z EILER , M. D.: ADADELTA: an adaptive learning rate method . arXiv preprint arXiv:1212.5701, 2012

2012 arXiv
[17]

K INGMA , D. P. and J. B A: Adam: A method for stochastic optimization . arXiv preprint arXiv:1412.6980, 2014

2014 arXiv
[18]

M ILDE , B. and A. K ÖHN : Open Source Automatic Speech Recognition for German . In Proc. of ITG. 2018

2018
[19]

C HEN , D

P ANAYOTOV, V., G. C HEN , D. P OVEY , and S. K HUDANPUR : Librispeech: an ASR cor- pus based on public domain audio books . In Proc. of IEEE ICASSP . 2015

2015
[20]

G ODFREY , J. J., E. C. H OLLIMAN , and J. M CDANIEL : SWITCHBOARD: Telephone speech corpus for research and development . In Proc. of IEEE ICASSP . 1992

1992
[21]

N GUYEN , S

H ERNANDEZ , F., V. N GUYEN , S. G HANNAY , N. T OMASHENKO , and Y. E STEVE : TED- LIUM 3: twice as much data and corpus repartition for experim ents on speaker adapta- tion. arXiv preprint arXiv:1805.04699, 2018

2018 arXiv
[22]

Language Resources and Evaluation , 2007

C ARLETTA , J.: Unleashing the killer corpus: experiences in creating the multi-everything AMI Meeting Corpus . Language Resources and Evaluation , 2007

2007
[23]

P AUL, D. B. and J. M. B AKER : The design for the Wall Street Journal-based CSR corpus. In Proc. of the workshop on Speech and Natural Language . 1992

1992
[24]

M ILDE , A

R ADECK -A RNETH , S., B. M ILDE , A. L ANGE , E. G OUVÊA , S. R ADOMSKI , M. M ÜHLHÄUSER , and C. B IEMANN : Open source german distant speech recognition: Corpus and acoustic model . In Text, Speech, and Dialogue. 2015

2015
[25]

S TEGEN , and T

K ÖHN , A., F. S TEGEN , and T. B AUMANN : Mining the Spoken Wikipedia for Speech Data and Beyond. In Proc. of LREC. 2016

2016
[26]

Springer Sci- ence & Business Media, 2013

W AHLSTER , W.: V erbmobil: foundations of speech-to-speech translation . Springer Sci- ence & Business Media, 2013

2013
[27]

B URGER , S. and F. S CHIEL : RVG 1 – A Database for Regional Variants of Contemporary German. In Proc. of LREC. 1998

1998
[28]

H ESS , W. J., K. J. K OHLER , and H.-G. T ILLMANN : The Phondat-verbmobil speech corpus. In F ourth European Conference on Speech Communication and Technology. 1995

1995
[29]

C HAN , W. and I. L ANE : Deep recurrent neural networks for acoustic modelling . arXiv preprint arXiv:1504.01482, 2015

2015 arXiv
[30]

H AN, K. J., A. C HANDRASHEKARAN , J. K IM, and I. L ANE : The CAPIO 2017 conver- sational speech recognition system. arXiv preprint arXiv:1801.00059, 2017

2017 arXiv
[31]

L ANGE , R

G AIDA , C., P. L ANGE , R. P ETRICK , P. P ROBA, A. M ALATAWY, and D. S UENDERMANN -O EFT : Comparing open-source speech recognition toolkits . Tech. Rep., DHBW Stuttgart, 2014

2014
[32]

B UI, and W

D ERNONCOURT , F., T. B UI, and W. C HANG : A Framework for Speech Recognition Benchmarking. Proc. of Interspeech, 2018

2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

G HOSHAL , G

P OVEY, D., A. G HOSHAL , G. B OULIANNE , L. B URGET , O. G LEMBEK , N. G OEL , M. H ANNEMANN , P. M OTLICEK , Y. Q IAN , P. S CHWARZ ET AL .: The Kaldi speech recognition toolkit. In Proc. of ASRU. 2011

work page 2011

[2] [2]

H ORI , S

W ATANABE , S., T. H ORI , S. K ARITA , T. H AYASHI, J. N ISHITOBA , Y. U NNO , N. E. Y. SOPLIN , J. H EYMANN , M. W IESNER , N. C HEN ET AL .: ESPnet: End-to-End Speech Processing Toolkit. arXiv preprint arXiv:1804.00015, 2018

arXiv 2018

[3] [3]

H ANAZAWA, G

W AIBEL , A., T. H ANAZAWA, G. H INTON , K. S HIKANO , and K. J. L ANG : Phoneme recognition using time-delay neural networks . In Readings in speech recognition . 1990

work page 1990

[4] [4]

M ANOHAR , D

G HAHREMANI , P., V. M ANOHAR , D. P OVEY , and S. K HUDANPUR : Acoustic Modelling from the Signal Domain Using CNNs . In Proc. of Interspeech. 2016

work page 2016

[5] [5]

IEEE Transactions on Information Theory , 1967

V ITERBI , A.: Error bounds for convolutional codes and an asymptotically optimum de- coding algorithm. IEEE Transactions on Information Theory , 1967

work page 1967

[6] [6]

C HO, and Y

B AHDANAU , D., K. C HO, and Y. B ENGIO : Neural machine translation by jointly learn- ing to align and translate . arXiv preprint arXiv:1409.0473, 2014

arXiv 2014

[7] [7]

F ERNÁNDEZ , F

G RAVES, A., S. F ERNÁNDEZ , F. G OMEZ , and J. S CHMIDHUBER : Connectionist tempo- ral classiﬁcation: labelling unsegmented sequence data wi th recurrent neural networks . In Proc. of ICML. 2006

work page 2006

[8] [8]

H ADDOW , and A

S ENNRICH , R., B. H ADDOW , and A. B IRCH : Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015

arXiv 2015