Two-stage Training for Chinese Dialect Recognition

Guofu Yang; Shugong Xu; Zongze Ren

REVIEW 4 major objections 5 minor 39 references

Two-stage Training for Chinese Dialect Recognition

T0 review · 4 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 A two-stage system that first learns phoneme sequences with CTC, then freezes those features to train a small BLSTM, reaches 88.88 percent accuracy on ten Chinese dialects.

desk verdict A plausible competition-winning recipe, but the key 10-point gain is confounded: the baseline lacks the ResNet14 front-end, so the paper overclaims the two-stage benefit without an architecture-matched control. read the letter →

arxiv 1908.02284 v2 pith:DHLA67NC submitted 2019-08-06 cs.CL cs.LGeess.AS

Zongze Ren , Guofu Yang , Shugong Xu This is my paper · ORCID

classification cs.CLcs.LGeess.AS

keywords Chinesedialectrecognitionlanguageidentificationtwo-stagetrainingconnectionisttemporalclassificationResNet14BLSTMintermediatefeaturesacousticmodel

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 argues that Chinese dialect identification improves when a model first learns to recognize phonemes and only then learns to classify dialects. The authors build a two-stage system: a ResNet14 with a two-layer BLSTM is trained with connectionist temporal classification to transcribe each utterance's phoneme sequence, and the trained features are frozen and fed to a second BLSTM that predicts one of ten dialects. On their test set this two-stage pipeline reaches 88.88 percent overall accuracy, compared with 78.85 percent for a one-stage BLSTM baseline and 87.24 percent for a three-stage variant, while using less training time than the three-stage variant. The concrete win is that phoneme-level supervision transfers to dialect identity, so a lightweight second network can do the classification.

What carries the argument

The load-bearing object is the intermediate feature map produced by the ResNet14 portion of the first-stage CTC acoustic model. This compact 5.36-million-parameter CNN is derived from ResNet-18 by reducing the residual blocks to six, and its role is to compress 40-dimensional log-mel filterbank frames into 512-dimensional frame-level representations that are then passed to the second-stage BLSTM. The CTC training signal is what gives these features their transferable content: it teaches the network to align its outputs to phoneme sequences without frame-level labels, and the paper's stated assumption is that the resulting high-dimensional features encode pronunciation and language-category information useful for dialect discrimination. The second stage then applies average pooling over time to convert frame-level posteriors into an utterance-level dialect decision.

What would settle it

Train the same second-stage BLSTM on intermediate features from an untrained, randomly initialized ResNet14, keeping everything else fixed; if accuracy stays near 88.88 percent, the first-stage CTC training is not the source of the gain. Alternatively, replace the CTC first-stage with a first-stage trained on a different label set, such as speaker identity, and look for a change in second-stage dialect accuracy.

Watch

Extended reading notes

Core claim

The central claim is a specific training recipe for spoken-language identification: first train an acoustic model end-to-end with CTC on phoneme-sequence annotations, freeze it, and then train a second two-layer BLSTM on the intermediate features of the frozen ResNet14 to output dialect posteriors under cross-entropy loss. The authors report that this two-stage system reaches 88.88 percent accuracy over ten Chinese dialects (87.72 percent for utterances at most three seconds and 90.04 percent for longer ones), beating a one-stage BLSTM baseline (78.85 percent) and a three-stage system that first aligns phonemes frame-by-frame (87.24 percent). They attribute the gain to a step-by-step schedule: backpropagating through the shared layers during the second stage causes the network to lose acoustic knowledge, so they freeze the first-stage network and train only the RNN. The paper also claims that BLSTM outperforms BGRU in this setup and that adding more BLSTM nodes does not help.

Load-bearing premise

The load-bearing premise is that features from a network trained only to predict phoneme sequences preserve enough dialect-distinguishing information; the paper states this assumption in Section 3.3 but does not ablate the first stage to test it.

Editorial extensions

If this is right

Because the first stage is trained once and frozen, reusing it makes per-dialect or per-language classifiers cheap to train: the second-stage RNN converges in six epochs.
The two-stage system beats the three-stage alignment-based system on overall, short, and long-utterance accuracy while using fewer total epochs, so extra frame-level alignment stages do not pay for themselves in this setting.
The largest relative improvement over the baseline is on short utterances, where accuracy rises from 77.60 percent to 87.72 percent, the condition where utterance-level language identification is usually hardest.
Using the same ResNet14 and BLSTM, replacing CTC phoneme supervision with direct dialect classification gives only 78.85 percent, so the ordering of tasks matters: phoneme knowledge first, dialect label second.

Reading between the lines

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

An untested extension is that the recipe may transfer to other language families: if a CTC-trained recognizer exists for a language group, its frozen intermediate features could seed a lightweight dialect or language classifier without retraining the whole acoustic model.
The paper's explanation that backpropagating shared layers makes the network forget acoustic knowledge is a testable design principle; a multi-task version of the same network trained jointly on phoneme CTC and dialect cross-entropy would directly show whether sequential freezing is necessary.
Because the first stage uses CTC rather than forced alignment, the two-stage idea suggests that unaligned audio-plus-text data, not just aligned speech corpora, may suffice to build dialect recognition systems.

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

You should know two things about this one. First, it is a competition system paper, and it is honest about that: the two-stage pipeline (CTC-trained ResNet14 intermediate features into a 2-layer BLSTM) is a practical recipe, and the two-vs-three-stage ablation is a genuinely useful control. Second, the headline result is not as clean as it looks. The one-stage baseline is a 2-layer BLSTM on log-mel features, with no ResNet14. The two- and three-stage systems both have ResNet14. So the 10-point gap conflates extra CNN capacity with the two-stage training scheme. The paper never trains a one-stage ResNet14+BLSTM directly on dialect labels. That missing baseline is load-bearing, and the stress-test note is right to put it first.

What is actually new is modest: applying CTC-trained acoustic-model features as inputs to a backend LID classifier is a variation on the bottleneck-feature line ([13-16]), and the CNN-BLSTM front end is close to [12]. The specific two-stage configuration on this 10-dialect Xunfei corpus is new, and the three-stage comparison seems not to have been published before. The internal numbers are consistent: two-stage beats three-stage under both durations, and the confusion-matrix analysis connects errors to dialect geography, which is a nice touch.

Where the soft spots are, in order of severity. The baseline mismatch is the big one; without a same-architecture one-stage control, the paper's Section 3.3 claim that step-by-step training preserves phonetic information is not actually demonstrated. Second, there are no error bars, no significance tests, and no held-out dev set description; the performance differences could be within run-to-run noise for all we know. Third, hyperparameters like learning rates and weight decay are described only as "different," not specified, and no code or data are released, so the result is not independently reproducible. Fourth, the abstract and conclusion call this state-of-the-art, which is unsubstantiated against published LID systems on standard benchmarks.

The transfer assumption the reader flagged is real but secondary; it is a mechanism question, and the missing architecture-matched baseline is the more direct experimental flaw.

Who should read this: anyone building practical LID systems for Chinese dialects or wanting a compact example of feature-transfer from CTC acoustic models. It is not a methods paper and adds no theoretical insight. It deserves a serious referee because the empirical recipe is plausible and the two-vs-three-stage result is interesting, but it would need substantial revision: add an architecture-matched baseline, report variance, and temper the claims. I would send it to review, and I would tell the authors to address the confound before I would cite the 10-point gain.

Referee Report

4 major / 5 minor

Summary. The paper proposes a two-stage training scheme for Chinese dialect identification (LID). In the first stage, a ResNet14 followed by a BLSTM is trained with a CTC loss to recognize phoneme sequences from log-mel filterbank features. The intermediate ResNet14 features are then frozen and fed, in the second stage, to a separate 2-layer BLSTM trained with cross-entropy loss to classify the utterance into one of ten dialects. The authors also explore a three-stage variant in which an additional ResNet14 AM is trained with frame-wise phoneme labels derived from the first-stage CTC alignments, and the intermediate features from that AM are then used for LID. On the Xunfei Chinese Dialect Recognition Challenge test set, the two-stage system achieves 88.88% overall accuracy, outperforming a one-stage baseline (78.85%) and the three-stage variant (87.24%), with shorter total training time. The paper reports that the system won first place among 110 challenge teams.

Significance. If the reported gains are attributable to the proposed two-stage training strategy, the paper would offer a practical recipe for exploiting phonetic annotations to improve LID in close dialects, and it provides a useful comparison of two-stage versus three-stage pipelines. The ablation of two versus three stages is a reasonable control, and the use of a frozen feature extractor in the second stage avoids end-to-end fine-tuning complexity. However, the significance is currently limited because the main comparison is not architecture-matched: the baseline lacks the ResNet14 front-end entirely, so the 10-point improvement confounds the training strategy with added model capacity. The paper also provides no error bars, significance testing, or specification of several key hyperparameters, making the headline accuracy difficult to interpret as evidence for the proposed mechanism.

major comments (4)

[§4.2 and §5.1, Table 2]
[§3.3]
[§5.2, Tables 3–4]
[§4.2]

minor comments (5)

[§3.2, Eq. (1)]
[Table 1]
[Figure 3]
[§1, first paragraph]
[§5.1, first paragraph]

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the two-stage pipeline is an empirical system evaluated on an external challenge test set, with no equation, fitted parameter, or self-citation that reduces to its own inputs.

full rationale

The paper reports an empirical two-stage LID system: stage one trains a ResNet14+RNN acoustic model with CTC on phoneme-sequence annotations, and stage two trains a 2-layer BLSTM on the frozen ResNet14 intermediate features using cross-entropy on dialect labels. The final accuracy numbers (Table 2) are measured on the Xunfei challenge test set, so the central result is externally anchored rather than derived from the training objective. No fitted parameter is renamed as a prediction: the intermediate features are not optimized toward dialect labels in stage one, and stage two is trained and then evaluated on held-out data. The stated assumption in Section 3.3 that high-dimensional features 'will contain information of pronunciation and language category' is a testable hypothesis, not a definitional equivalence, and the paper does not use it as a proof step. There are no author self-citations carrying load-bearing weight, no uniqueness theorem imported from prior work, and no ansatz smuggled in via citation; the architecture is described directly in Section 3.1 and Table 1. The clearest weakness is that the baseline is not architecture-matched: Section 4.2 says the baseline is 'a one-stage RNN system, the RNN structure is the same as the last stage containing 2-layer BLSTM,' whereas both multi-stage systems add the ResNet14 front-end, so the 10-point gain conflates architectural capacity with the two-stage training procedure. That confound is a correctness/experimental-design concern, not circularity, because the comparison does not define the result into existence nor fit the reported accuracy. Accordingly, the circularity score is 0.

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

The central result depends on the choice of hyperparameters (BLSTM size, dropout, learning rates) and on the assumption that CTC-trained features transfer to dialect classification. No new entities are invented. The main unstated cost is the private dataset and the absence of released code.

free parameters (4)

BLSTM hidden size = 256
Chosen by hand; 384 and 512 tested, with 256 giving best results for the two-stage BLSTM.
Number of BLSTM layers = 2
Chosen by hand; 3 layers did not improve accuracy.
Dropout rate in LID stage = 0.5
Set for the second/third stage LID network.
Learning rate and weight decay per stage = not reported
The paper says 'different learning rates and weight decay in different stages' but does not give values; these affect convergence and final accuracy.

assumptions (4)

standard math CTC loss is a valid objective for training the acoustic model to predict phoneme sequences without alignment.
Paper relies on Graves et al. [17] for CTC; unproved in this paper.
domain assumption The provided phonetic sequence annotations and dialect labels are correct.
The training and evaluation rely on challenge data annotations; no quality check is reported.
ad hoc to paper Intermediate features from a CTC-trained ResNet14 contain pronunciation and language category information useful for LID.
Stated as an assumption in Section 3.3 ('we assume that the high-dim features... will contain information of pronunciation and language category'). This is the load-bearing transfer assumption.
domain assumption Training stages sequentially rather than jointly prevents the network from 'forgetting' acoustic knowledge.
Empirical claim in Section 5.2 and Section 3.3, not proven theoretically.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Two-stage Training for Chinese Dialect Recognition." pith.science (2026). https://pith.science/paper/DHLA67NC

@misc{pith2026190802284,
  author       = {Pith},
  title        = {Pith review of: Two-stage Training for Chinese Dialect Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DHLA67NC}},
  note         = {Machine review of arXiv:1908.02284}
}

read the original abstract

In this paper, we present a two-stage language identification (LID) system based on a shallow ResNet14 followed by a simple 2-layer recurrent neural network (RNN) architecture, which was used for Xunfei (iFlyTek) Chinese Dialect Recognition Challenge and won the first place among 110 teams. The system trains an acoustic model (AM) firstly with connectionist temporal classification (CTC) to recognize the given phonetic sequence annotation and then train another RNN to classify dialect category by utilizing the intermediate features as inputs from the AM. Compared with a three-stage system we further explore, our results show that the two-stage system can achieve high accuracy for Chinese dialects recognition under both short utterance and long utterance conditions with less training time.

Figures

Figures reproduced from arXiv: 1908.02284 by the authors.

**Figure 1.** Training architecture of two-stage system where X is the utterance and p is a CTC path. Then the network can be trained to optimize the CTC function −lnP(z|X) by the given sequence labeling. For the LID task, we use the multiclass cross-entropy loss for classification: LCE = − XT j=1 yilogPj (2) where yi is the ground truth label and Pj is the output probability distribution. 3.3. Two-stage system [PITH_FULL_IMAG… view at source ↗

**Figure 3.** Data distribution of time The baseline we use for comparison is a one-stage RNN system, the RNN structure is the same as the last stage containing 2-layer BLSTM and directly trained to recognize dialect category. In the process of evaluation, we compute the accuracy of the two sub-tasks and the whole test set to evaluate the performance of each system. 5. Results 5.1. Comparison of different stage systems First of… view at source ↗

**Figure 4.** Comparison of confusion matrices produced by the two-stage system (left) and the three-stage system (right) does not work, but adding another layer makes sense in shortduration task. Moreover, with the same layers and nodes, BLSTM outperforms BGRU in the two sub-tasks. We believe that sound related tasks do not need a very deep network as image related tasks, that is also the reason why we use a shallow ResNet14 a… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 34 canonical work pages

[12]

Multi- lingual bottleneck features for language recognition,

R. F ´er, P. Matˇejka, F. Gr´ezl, O. Plchot, and J. ˇCernock`y, “Multi- lingual bottleneck features for language recognition,” inSixteenth Annual Conference of the International Speech Communication Association, 2015

work page 2015
[1]

The task intro- duced in this paper is more challenging than general LID tasks cause we use a dialect database which contains 10 dialects in China

Introduction The aim of language identiﬁcation (LID) is to determine the lan- guage of an utterance and can be deﬁned as a variable-length se- quence classiﬁcation task on the utterance-level. The task intro- duced in this paper is more challenging than general LID tasks cause we use a dialect database which contains 10 dialects in China. The dialects’ re...

work page
[2]

Our system won the ﬁrst place in Xunfei (iFlyTek) Chi- nese Dialect Recognition Challenge, which attracted 110 teams joining in

We explore a new Chinese dialect database by a two- stage LID system and achieve high accuracy of 89%. Our system won the ﬁrst place in Xunfei (iFlyTek) Chi- nese Dialect Recognition Challenge, which attracted 110 teams joining in

work page
[3]

Compared with other methods which recognize physical states or triphones, we train an AM end-to-end and use CTC [17] to recognize the phoneme sequence of an ut- terance directly

work page
[4]

We train an AM in the ﬁrst stage and then use the intermediate features from the AM as inputs to train an RNN to compute posteriors for LID in the second stage

We use a specially designed ResNet14 followed by an RNN instead of using fully connected layers to develop the two-stage LID system. We train an AM in the ﬁrst stage and then use the intermediate features from the AM as inputs to train an RNN to compute posteriors for LID in the second stage

work page
[5]

The results show that the per- formance is slightly worse than the two-stage system

We further investigate a three-stage system, we ﬁrst train an AM to align the phoneme label and then train an- other AM to predict the phoneme of each frame, ﬁnally we use the intermediate features from the second-stage’s network to train an RNN. The results show that the per- formance is slightly worse than the two-stage system. The remainder of the pape...

work page
[6]

Related works ASR [18] task enables the recognition and translation of spo- ken language into text. Traditionally, we can train an AM based on frame-wise cross-entropy loss to recognize phoneme, which requires tedious label alignment procedure such as Hidden Markov Model and Gaussian Mixture Model (HMM-GMM) paradigm. Then we can use a pronunciation model ...

work page Pith review arXiv 1908
[7]

Network structure The major network structure we use in the two-stage system can be divided to the CNN part and the RNN part, as described in Table 1

Two-stage system overview 3.1. Network structure The major network structure we use in the two-stage system can be divided to the CNN part and the RNN part, as described in Table 1. Given the input data of shape T× 40, where T is the frame length of an utterance, we ﬁnally get 512-dimensional frame-level representation and N is the number of phonemes or d...

work page

Show all 39 references

[8]

Experiments 4.1. Data description We use a database covering 10 most widespread Chinese di- alects, the dialects are Ningxia, Hefei, Sichuan, Shanxi, Chang- sha, Hebei, Nanchang, Shanghai, Kekka and Fujian. Each di- alect has 6-hour audio data. For the training set, there will...
[9]

Results 5.1. Comparison of different stage systems First of all, we compare the two-stage system and the three- stage system trained with phonetic sequence annotation and dialect category label with the baseline trained only with di- alect category label. The two multi-stage s...
[10]

The system links the different stages by using inter- mediate features extracted by a shallow ResNet14 architecture

Conclusions In this work, we propose an acoustic model based on ResNet14 followed by an RNN to recognize phoneme sequence directly with CTC loss and train a simple RNN lastly to get posteri- ors for recognizing dialect category, forming a two-stage LID system. The system links...
[11]

Deep neural network approaches to speaker and language recognition,

F. Richardson, D. Reynolds, and N. Dehak, “Deep neural network approaches to speaker and language recognition,” IEEE Signal Processing Letters, vol. 22, no. 10, pp. 1671–1675, 2015

2015
[13]

Neural network bottleneck features for language identiﬁcation,

P. Matejka, L. Zhang, T. Ng, H. S. Mallidi, O. Glembek, J. Ma, and B. Zhang, “Neural network bottleneck features for language identiﬁcation,” in Proceedings of Odyssey , vol. 2014, 2014, pp. 299–304

2014
[14]

Deep bottleneck features for spoken language identiﬁca- tion,

B. Jiang, Y . Song, S. Wei, J.-H. Liu, I. V . McLoughlin, and L.- R. Dai, “Deep bottleneck features for spoken language identiﬁca- tion,” PloS one, vol. 9, no. 7, p. e100795, 2014

2014
[15]

Automatic lan- guage identiﬁcation using deep neural networks,

I. Lopez-Moreno, J. Gonzalez-Dominguez, O. Plchot, D. Mar- tinez, J. Gonzalez-Rodriguez, and P. Moreno, “Automatic lan- guage identiﬁcation using deep neural networks,” in IEEE Inter- national Conference on Acoustics, 2014

2014
[16]

An end-to-end approach to language identiﬁcation in short utterances using convolutional neural networks,

A. Lozano-Diez, R. Zazo-Candil, J. Gonzalez-Dominguez, D. T. Toledano, and J. Gonzalez-Rodriguez, “An end-to-end approach to language identiﬁcation in short utterances using convolutional neural networks,” in Sixteenth Annual Conference of the Interna- tional Speech Communicat...

2015
[17]

End-to- end language identiﬁcation using high-order utterance representa- tion with bilinear pooling,

M. Jin, Y . Song, I. Mcloughlin, W. Guo, and L. R. Dai, “End-to- end language identiﬁcation using high-order utterance representa- tion with bilinear pooling,” in Interspeech 2017, 2017

2017
[18]

Stacked long-term tdnn for spoken language recognition

D. Garcia-Romero and A. McCree, “Stacked long-term tdnn for spoken language recognition.” in INTERSPEECH, 2016, pp. 3226–3230

2016
[19]

Automatic language identiﬁcation using long short-term memory recurrent neural networks,

J. Gonzalez-Dominguez, I. Lopez-Moreno, H. Sak, J. Gonzalez- Rodriguez, and P. J. Moreno, “Automatic language identiﬁcation using long short-term memory recurrent neural networks,” in Fif- teenth Annual Conference of the International Speech Communi- cation Association, 2014

2014
[20]

End-to-end language identiﬁcation using attention-based recurrent neural net- works,

W. Geng, W. Wang, Y . Zhao, X. Cai, and B. Xu, “End-to-end language identiﬁcation using attention-based recurrent neural net- works,” Interspeech 2016, pp. 2944–2948, 2016

2016
[21]

Spoken language identiﬁcation us- ing lstm-based angular proximity

G. Gelly and J.-L. Gauvain, “Spoken language identiﬁcation us- ing lstm-based angular proximity.” in INTERSPEECH, 2017, pp. 2566–2570

2017
[22]

Utterance-level end-to-end language identiﬁcation using attention-based cnn-blstm,

W. Cai, D. Cai, S. Huang, and M. Li, “Utterance-level end-to-end language identiﬁcation using attention-based cnn-blstm,” arXiv preprint arXiv:1902.07374, 2019

1902 arXiv
[23]

Dnn based embeddings for language recognition,

A. Lozano-Diez, O. Plchot, P. Matejka, and J. Gonzalez- Rodriguez, “Dnn based embeddings for language recognition,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 5184–5188

2018
[24]

Phonetic tempo- ral neural model for language identiﬁcation,

Z. Tang, W. Dong, Y . Chen, L. Li, and A. Abel, “Phonetic tempo- ral neural model for language identiﬁcation,” IEEE/ACM Trans- actions on Audio Speech & Language Processing , vol. 26, no. 1, pp. 134–144, 2017

2017
[25]

Parallel pho- netically aware dnns and lstm-rnns for frame-by-frame discrimi- native modeling of spoken language identiﬁcation,

R. Masumura, T. Asami, H. Masataki, and Y . Aono, “Parallel pho- netically aware dnns and lstm-rnns for frame-by-frame discrimi- native modeling of spoken language identiﬁcation,” in IEEE In- ternational Conference on Acoustics, 2017

2017
[26]

Using deep neural networks for identiﬁcation of slavic languages from acoustic signal,

L. Mateju, P. Cerva, J. Zdansky, and R. Safarik, “Using deep neural networks for identiﬁcation of slavic languages from acoustic signal,” in Proc. Interspeech 2018 , 2018, pp. 1803–1807. [Online]. Available: http://dx.doi.org/10.21437/ Interspeech.2018-1165

2018
[27]

Con- nectionist temporal classiﬁcation: labelling unsegmented se- quence data with recurrent neural networks,

A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classiﬁcation: labelling unsegmented se- quence data with recurrent neural networks,” in Proceedings of the 23rd international conference on Machine learning . ACM, 2006, pp. 369–376

2006
[28]

Deep neural networks for acoustic modeling in speech recognition,

G. Hinton, L. Deng, D. Yu, G. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V . Vanhoucke, P. Nguyen, B. Kingsburyet al., “Deep neural networks for acoustic modeling in speech recognition,” IEEE Signal processing magazine, vol. 29, 2012

2012
[29]

Speech recognition with deep recurrent neural networks,

A. Graves, A.-r. Mohamed, and G. Hinton, “Speech recognition with deep recurrent neural networks,” in2013 IEEE international conference on acoustics, speech and signal processing . IEEE, 2013, pp. 6645–6649

2013
[30]

Towards end-to-end speech recognition with recurrent neural networks,

A. Graves and N. Jaitly, “Towards end-to-end speech recognition with recurrent neural networks,” in International Conference on Machine Learning, 2014, pp. 1764–1772

2014
[31]

Deep speech 2: End-to-end speech recognition in english and mandarin,

D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Bat- tenberg, C. Case, J. Casper, B. Catanzaro, Q. Cheng, G. Chen et al., “Deep speech 2: End-to-end speech recognition in english and mandarin,” in International conference on machine learning , 2016, pp. 173–182

2016
[32]

An end-to-end trainable neural net- work for image-based sequence recognition and its application to scene text recognition,

B. Shi, X. Bai, and C. Yao, “An end-to-end trainable neural net- work for image-based sequence recognition and its application to scene text recognition,” IEEE transactions on pattern analysis and machine intelligence, vol. 39, no. 11, pp. 2298–2304, 2017

2017
[33]

Towards end-to-end speech recognition with deep convolutional neural networks,

Y . Zhang, M. Pezeshki, P. Brakel, S. Zhang, C. Laurent, Y . Ben- gio, and A. Courville, “Towards end-to-end speech recognition with deep convolutional neural networks,” Interspeech 2016, pp. 410–414, 2016

2016
[34]

Residual convolutional ctc networks for automatic speech recognition,

Y . Wang, X. Deng, S. Pu, and Z. Huang, “Residual convolutional ctc networks for automatic speech recognition,” arXiv preprint arXiv:1702.07793, 2017

2017 arXiv
[35]

Residual lstm: Design of a deep recurrent architecture for distant speech recognition,

J. Kim, M. El-Khamy, and J. Lee, “Residual lstm: Design of a deep recurrent architecture for distant speech recognition,” Proc. Interspeech 2017, pp. 1591–1595, 2017

2017
[36]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

2016
[37]

Bidirectional recurrent neu- ral networks,

M. Schuster and K. K. Paliwal, “Bidirectional recurrent neu- ral networks,” IEEE Transactions on Signal Processing , vol. 45, no. 11, pp. 2673–2681, 1997

1997
[38]

Speaker embedding extraction with phonetic information,

Y . Liu, L. He, J. Liu, and M. T. Johnson, “Speaker embedding extraction with phonetic information,” in Proc. Interspeech 2018 , 2018, pp. 2247–2251. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2018-1226

2018 doi
[39]

Training multi-task adversarial network for extracting noise-robust speaker embedding,

J. Zhou, T. Jiang, L. Li, Q. Hong, Z. Wang, and B. Xia, “Training multi-task adversarial network for extracting noise-robust speaker embedding,” arXiv preprint arXiv:1811.09355, 2018

2018 arXiv

Pith tools

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

[1] [12]

Multi- lingual bottleneck features for language recognition,

R. F ´er, P. Matˇejka, F. Gr´ezl, O. Plchot, and J. ˇCernock`y, “Multi- lingual bottleneck features for language recognition,” inSixteenth Annual Conference of the International Speech Communication Association, 2015

work page 2015

[2] [1]

The task intro- duced in this paper is more challenging than general LID tasks cause we use a dialect database which contains 10 dialects in China

Introduction The aim of language identiﬁcation (LID) is to determine the lan- guage of an utterance and can be deﬁned as a variable-length se- quence classiﬁcation task on the utterance-level. The task intro- duced in this paper is more challenging than general LID tasks cause we use a dialect database which contains 10 dialects in China. The dialects’ re...

work page

[3] [2]

Our system won the ﬁrst place in Xunfei (iFlyTek) Chi- nese Dialect Recognition Challenge, which attracted 110 teams joining in

We explore a new Chinese dialect database by a two- stage LID system and achieve high accuracy of 89%. Our system won the ﬁrst place in Xunfei (iFlyTek) Chi- nese Dialect Recognition Challenge, which attracted 110 teams joining in

work page

[4] [3]

Compared with other methods which recognize physical states or triphones, we train an AM end-to-end and use CTC [17] to recognize the phoneme sequence of an ut- terance directly

work page

[5] [4]

We train an AM in the ﬁrst stage and then use the intermediate features from the AM as inputs to train an RNN to compute posteriors for LID in the second stage

We use a specially designed ResNet14 followed by an RNN instead of using fully connected layers to develop the two-stage LID system. We train an AM in the ﬁrst stage and then use the intermediate features from the AM as inputs to train an RNN to compute posteriors for LID in the second stage

work page

[6] [5]

The results show that the per- formance is slightly worse than the two-stage system

We further investigate a three-stage system, we ﬁrst train an AM to align the phoneme label and then train an- other AM to predict the phoneme of each frame, ﬁnally we use the intermediate features from the second-stage’s network to train an RNN. The results show that the per- formance is slightly worse than the two-stage system. The remainder of the pape...

work page

[7] [6]

Related works ASR [18] task enables the recognition and translation of spo- ken language into text. Traditionally, we can train an AM based on frame-wise cross-entropy loss to recognize phoneme, which requires tedious label alignment procedure such as Hidden Markov Model and Gaussian Mixture Model (HMM-GMM) paradigm. Then we can use a pronunciation model ...

work page Pith review arXiv 1908

[8] [7]

Network structure The major network structure we use in the two-stage system can be divided to the CNN part and the RNN part, as described in Table 1

Two-stage system overview 3.1. Network structure The major network structure we use in the two-stage system can be divided to the CNN part and the RNN part, as described in Table 1. Given the input data of shape T× 40, where T is the frame length of an utterance, we ﬁnally get 512-dimensional frame-level representation and N is the number of phonemes or d...

work page