Two-Stage Data Augmentation for Low-Resourced Speech Recognition

Hartmann, William M.; Ng, Tim; Hsiao, Roger; Tsakalidis, Stavros; Schwartz, Richard

doi:10.21437/interspeech.2016-1386

Cited by 37 publications

(16 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To address the lack of training data, however, a number of modifications have been made to the standard pipeline. These approaches include data augmentation [23,5,14]; the use of web data [20]; extensive system combination [31]; and the use of multiple languages [4,24]. Furthermore the concept of low resource can be applied beyond the availability of training data to include linguistic resources such as an accurate lexicon.…”

Section: Low-resource Speech Recognitionmentioning

confidence: 99%

Low-Resource Speech Recognition and Keyword-Spotting

Gales

Knill

Ragni

2017

Speech and Computer

View full text Add to dashboard Cite

The IARPA Babel program ran from March 2012 to November 2016. The aim of the program was to develop agile and robust speech technology that can be rapidly applied to any human language in order to provide effective search capability on large quantities of real world data. This paper will describe some of the developments in speech recognition and keyword-spotting during the lifetime of the project. Two technical areas will be briefly discussed with a focus on techniques developed at Cambridge University: the application of deep learning for low-resource speech recognition; and efficient approaches for keyword spotting. Finally a brief analysis of the Babel speech language characteristics and language performance will be presented.

show abstract

Section: Low-resource Speech Recognitionmentioning

confidence: 99%

Low-Resource Speech Recognition and Keyword-Spotting

Gales

Knill

Ragni

2017

Speech and Computer

View full text Add to dashboard Cite

show abstract

“…In [2] data augmentation by creating multiple versions of the original signal with various speed factors has been shown to improve ASR performance across various tasks. This approach is used more elaborately in [3] where in addition to data augmentation at the signal level with noise addition and speed perturbation, data is augmented in a second stage, at the feature level with a fMLLR-based technique applied to bottleneck features. Similarly in [4], data augmentation is performed at the feature level via vocal tract length perturbation (VTLP) and stochastic feature mapping (SFM), to improve ASR performance in low resource settings.…”

Section: Related Workmentioning

confidence: 99%

“…Following an approach introduced in Ko et al [2] and used by Hartmann et al [3], we perform speed perturbation to generate modified copies of the source audio. This approach modifies the speed of each file by a multiplicative factor drawn at random uniformly between 0.9 and 1.1.…”

Section: Speed Modificationmentioning

confidence: 99%

Data Augmentation Improves Recognition of Foreign Accented Speech

et al. 2018

View full text Add to dashboard Cite

Speech recognition of foreign accented (non-native or L2) speech remains a challenge to the state-of-the-art. The most common approach to address this scenario involves the collection and transcription of accented speech, and incorporating this into the training data. However, the amount of accented data is dwarfed by the amount of material from native (L1) speakers, limiting the impact of the additional material. In this work, we address this problem via data augmentation. We create modified copies of two accents, Latin American and Asian accented English speech with voice transformation (modifying glottal source and vocal tract parameters), noise addition, and speed modification. We investigate both supervised (where transcription of the accented data is available) and unsupervised approaches to using the accented data and associated augmentations. We find that all augmentations provide improvements, with the largest gains coming from speed modification, then voice transformation and noise addition providing the least improvement. The improvements from training accent specific models with the augmented data are substantial. Improvements from supervised and unsupervised adaptation (or training with soft labels) with the augmented data are relatively minor. Overall, we find speed modification to be a remarkably reliable data augmentation technique for improving recognition of foreign accented speech. Our strategies with associated augmentations provide Word Error Rate (WER) reductions of up to 30% relative over a baseline trained with only the accented data.

show abstract

“…The amounts of their training and test data are listed in Table 1. For each language the training data is doubled by one copy of augmentation created with varying speeds and adding Babel noises, as described in [21]. In this paper we report STT system performance (in terms of WER -word error rate) measured on the test sets given in Table 1.…”

Section: Training Tdnn-hmm Hybrid Systemsmentioning

confidence: 99%

Improving Deliverable Speech-to-Text Systems with Multilingual Knowledge Transfer

Jeff¹,

Keith²,

Ng³

et al. 2017

Interspeech 2017

Self Cite

View full text Add to dashboard Cite

This paper reports our recent progress on using multilingual data for improving speech-to-text (STT) systems that can be easily delivered. We continued the work BBN conducted on the use of multilingual data for improving Babel evaluation systems, but focused on training time-delay neural network (TDNN) based chain models. As done for the Babel evaluations, we used multilingual data in two ways: first, to train multilingual deep neural networks (DNN) for extracting bottleneck (BN) features, and second, for initializing training on target languages. Our results show that TDNN chain models trained on multilingual DNN bottleneck features yield significant gains over their counterparts trained on MFCC plus i-vector features. By initializing from models trained on multilingual data, TDNN chain models can achieve great improvements over random initializations of the network weights on target languages. Two other important findings are: 1) initialization with multilingual TDNN chain models produces larger gains on target languages that have less training data; 2) inclusion of target languages in multilingual training for either BN feature extraction or initialization have limited impact on performance measured on the target languages. Our results also reveal that for TDNN chain models, the combination of multilingual BN features and multilingual initialization achieves the best performance on all target languages.

show abstract

Two-Stage Data Augmentation for Low-Resourced Speech Recognition

Cited by 37 publications

References 22 publications

Low-Resource Speech Recognition and Keyword-Spotting

Low-Resource Speech Recognition and Keyword-Spotting

Data Augmentation Improves Recognition of Foreign Accented Speech

Improving Deliverable Speech-to-Text Systems with Multilingual Knowledge Transfer

Contact Info

Product

Resources

About