An Investigation of Mixup Training Strategies for Acoustic Models in ASR

Medennikov, Ivan; Khokhlov, Yuri Y.; Romanenko, Aleksei; Popov, Dmitry; Tomashenko, Natalia; Sorokin, Ivan; Zatvornitskiy, Alexander

doi:10.21437/interspeech.2018-2191

Cited by 33 publications

(17 citation statements)

References 20 publications

(29 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…39phone and 84phone means the number of phones we used to training the model. Attention and mix-up [22] are the method when we trained the model. From the table 2, we can see that the performances of different networks are similar.…”

Section: Acoustic Modelsmentioning

confidence: 99%

The LeVoice Far-Field Speech Recognition System for VOiCES from a Distance Challenge 2019

Liang¹,

Yang²,

Wang³

et al. 2019

Interspeech 2019

View full text Add to dashboard Cite

This paper describes our submission to the "VOiCES from a Distance Challenge 2019", which is designed to foster research in the area of speaker recognition and automatic speech recognition (ASR) with a special focus on single channel distant/far-field audio under noisy conditions. We focused on the ASR task under a fixed condition in which the training data was clean and small, but the development data and test data were noisy and unmatched. Thus we developed the following major technical points for our system, which included data augmentation, weighted-prediction-error based speech enhancement, acoustic models based on different networks, TDNN or LSTM based language model rescore, and ROVER. Experiments on the development set and the evaluation set showed that the front-end processing, data augmentation and system fusion made the main contributions for the performance increasing, and the final word error rate results based on our system scored 15.91% and 19.6% respectively.

show abstract

Section: Acoustic Modelsmentioning

confidence: 99%

The LeVoice Far-Field Speech Recognition System for VOiCES from a Distance Challenge 2019

Liang¹,

Yang²,

Wang³

et al. 2019

Interspeech 2019

View full text Add to dashboard Cite

show abstract

“…This approach performs on-the-fly generation of virtual training examples by combining the existing ones. We used Kaldi-compatible implementation 2 described in our previous paper [29]. Figure 1 presents the preparation of 3 datasets D1, D2, and D3, used for training of the acoustic models.…”

Section: Signal Enhancement and Data Augmentationmentioning

confidence: 99%

“…• AM1: Time Delay Neural Network (TDNN) [33] trained with LF-MMI criterion [34] on D1 dataset and SDBN features. Training was performed with mixup data augmentation [29] and backstitch regularization [35].…”

Section: Final Acoustic Modelsmentioning

confidence: 99%

The STC System for the CHiME 2018 Challenge

Medennikov¹,

Sorokin²,

Romanenko³

et al. 2018

5th International Workshop on Speech Processing in Everyday Environments (CHiME 2018)

Self Cite

View full text Add to dashboard Cite

This paper describes the Speech Technology Center (STC) system for the 5th CHiME challenge. This challenge considers the problem of distant multi-microphone conversational speech recognition in everyday home environments. Our efforts were focused on the single-array track, however, we participated in the multiple-array track as well. The system is in the ranking A of the challenge: acoustic models remain frame-level tied phonetic targets, lexicon and language model are not changed compared to the conventional ASR baseline. Our system employs a combination of 4 acoustic models based on convolutional and recurrent neural networks. Speaker adaptation with target speaker masks and multi-channel speaker-aware acoustic model with neural network beamforming are two major features of the system. Moreover, various techniques for improving acoustic models are applied, including array synchronization, data cleanup, alignment transfer, mixup, speed perturbation data augmentation, room simulation, and backstitch training. Our system scored 3rd in the single-array track with Word Error Rate (WER) of 55.5% and 4th in the multiple-array track with WER of 55.6% on the evaluation data, achieving a substantial improvement over the baseline system.

show abstract

“…Weights for the combination are proportional to ξ α and (1 − ξ) α , where ξ and (1 − ξ) are the same weights that are applied for the corresponding feature vectors, and α is a scaling factor 1 . Other ways to apply mixup for ASR can be found in [39].…”

Section: Generate a New Synthetic Sequence Of Input Vectors Asmentioning

confidence: 99%

Speaker Adaptive Training and Mixup Regularization for Neural Network Acoustic Models in Automatic Speech Recognition

Tomashenko¹,

Khokhlov²,

Estève³

2018

Interspeech 2018

Self Cite

View full text Add to dashboard Cite

This work investigates speaker adaptation and regularization techniques for deep neural network acoustic models (AMs) in automatic speech recognition (ASR) systems. In previous works, GMM-derived (GMMD) features have been shown to be an efficient technique for neural network AM adaptation. In this paper, we propose and investigate a novel way to improve speaker adaptive training (SAT) for neural network AMs using GMMD features. The idea is based on using inaccurate transcriptions from ASR for adaptation during neural network training, while keeping the exact transcriptions for targets of neural networks. In addition, we apply a mixup technique, recently proposed for classification tasks, to acoustic models for ASR and investigate the impact of this technique on speaker adapted acoustic models. Experimental results on the TED-LIUM corpus show that the proposed approaches provide an additional gain in speech recognition performance in comparison with the speaker adapted AMs.

show abstract

An Investigation of Mixup Training Strategies for Acoustic Models in ASR

Cited by 33 publications

References 20 publications

The LeVoice Far-Field Speech Recognition System for VOiCES from a Distance Challenge 2019

The LeVoice Far-Field Speech Recognition System for VOiCES from a Distance Challenge 2019

The STC System for the CHiME 2018 Challenge

Speaker Adaptive Training and Mixup Regularization for Neural Network Acoustic Models in Automatic Speech Recognition

Contact Info

Product

Resources

About