Collecting and evaluating speech recognition corpora for 11 South African languages

Badenhorst, Jaco; Heerden, Charl Johannes van; Davel, Marelie H.; Barnard, Etienne

doi:10.1007/s10579-011-9152-1

Cited by 41 publications

(11 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This is especially useful for underresourced languages lacking a substantial set of training data for a complete ASR system (Badenhorst et al, 2011;Livescu et al, 2012;Li, 2008a, 2008b). The accuracy of this strategy depends closely on the fit between the phone set used to build the aligner and the phone set of the target language.…”

Section: A Forced Alignmentmentioning

confidence: 99%

“…Second, forced alignment systems are often trained on large corpora (often between ten and hundreds of hours of speech) from a large number of speakers, e.g., between 40 (Malfrère et al, 2003) and 630 speakers (Garofolo et al, 1993). Even an automatic speech recognition system (ASR) built for under-resourced languages typically involves recordings from hundreds of speakers (Badenhorst et al, 2011). Transcribed corpora from endangered language documentation projects typically come from a handful of speakers and typically consist of between 5-60 h of recordings.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Using automatic alignment to analyze endangered language data: Testing the viability of untrained alignment

DiCanio

Nam

Whalen

et al. 2013

The Journal of the Acoustical Society of America

View full text Add to dashboard Cite

While efforts to document endangered languages have steadily increased, the phonetic analysis of endangered language data remains a challenge. The transcription of large documentation corpora is, by itself, a tremendous feat. Yet, the process of segmentation remains a bottleneck for research with data of this kind. This paper examines whether a speech processing tool, forced alignment, can facilitate the segmentation task for small data sets, even when the target language differs from the training language. The authors also examined whether a phone set with contextualization outperforms a more general one. The accuracy of two forced aligners trained on English (HMALIGN and P2FA) was assessed using corpus data from Yolox ochitl Mixtec. Overall, agreement performance was relatively good, with accuracy at 70.9% within 30 ms for HMALIGN and 65.7% within 30 ms for P2FA. Segmental and tonal categories influenced accuracy as well. For instance, additional stop allophones in HMALIGN's phone set aided alignment accuracy. Agreement differences between aligners also corresponded closely with the types of data on which the aligners were trained. Overall, using existing alignment systems was found to have potential for making phonetic analysis of small corpora more efficient, with more allophonic phone sets providing better agreement than general ones.

show abstract

Section: A Forced Alignmentmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Using automatic alignment to analyze endangered language data: Testing the viability of untrained alignment

DiCanio

Nam

Whalen

et al. 2013

The Journal of the Acoustical Society of America

View full text Add to dashboard Cite

show abstract

“…Initial projects were funded by the Department of Arts, Culture, Science and Technology (DACST) and subsequently by the Departments of Arts and Culture (DAC) and Science and Technology (DST), respectively, after the two departments became separate entities. For instance, the African Speech Technology (AST) project [1] was supported by DACST, while DAC funded projects like Lwazi [2,3] and the National Centre for Human Language Technology (NCHLT) speech [4,5] and text [6] projects. The recently-established South African Centre for Digital Language Resources (https://www.sadilar.…”

Section: Introductionmentioning

confidence: 99%

“…In South Africa, the AST, Lwazi, and the first NCHLT project relied on data collection to create speech resources for the indigenous languages. During the Lwazi project, telephone speech was collected (between four and ten hours per language [2]), while the aim of the first NCHLT project was to collect 50-60 h of orthographically-transcribed, broadband speech in each of the country's 11 official languages [4].…”

Section: Introductionmentioning

confidence: 99%

The Usefulness of Imperfect Speech Data for ASR Development in Low-Resource Languages

Badenhorst¹,

Wet

2019

Information

Self Cite

View full text Add to dashboard Cite

When the National Centre for Human Language Technology (NCHLT) Speech corpus was released, it created various opportunities for speech technology development in the 11 official, but critically under-resourced, languages of South Africa. Since then, the substantial improvements in acoustic modeling that deep architectures achieved for well-resourced languages ushered in a new data requirement: their development requires hundreds of hours of speech. A suitable strategy for the enlargement of speech resources for the South African languages is therefore required. The first possibility was to look for data that has already been collected but has not been included in an existing corpus. Additional data was collected during the NCHLT project that was not included in the official corpus: it only contains a curated, but limited subset of the data. In this paper, we first analyze the additional resources that could be harvested from the auxiliary NCHLT data. We also measure the effect of this data on acoustic modeling. The analysis incorporates recent factorized time-delay neural networks (TDNN-F). These models significantly reduce phone error rates for all languages. In addition, data augmentation and cross-corpus validation experiments for a number of the datasets illustrate the utility of the auxiliary NCHLT data.

show abstract

“…The Lwazi project was set up to develop a telephone-based speech-driven information system in South Africa. In the project, the Lwazi ASR corpus [77] words, dates, and numbers. A 5,000-word pronunciation dictionary was also created for each language, which covers only the most common words in the language.…”

Section: The Lwazi Speech Corpusmentioning

confidence: 99%

Multi-task learning deep neural networks for automatic speech recognition

Chen¹

View full text Add to dashboard Cite

Chapter 1 Introduction 1 1.1 Why Multi-task Learning (MTL) for ASR? 3 1.2 How to Apply MTL to ASR? 5 1.3 Thesis Outline 6 Chapter 2 Review of Automatic Speech Recognition 8 2.1 Automatic Speech Recognition 8 2.2 Language Model 9 2.3 Acoustic Model 2.3.1 Hidden Markov Model (HMM) 2.3.2 Gaussian Mixture Model (GMM) 2.3.3 Deep Neural Network (DNN) 2.4 Phonetic Unit Selection 2.4.1 Context-independent (CI) Units 2.4.2 Context-dependent (CD) Units 2.5 Context-Dependent Acoustic Modeling vi 2.6 Data Scarcity Problem in ASR 2.7 ASR System Performance Evaluation Metric 2.8 Summary Chapter 3 MUlti-task Learning Deep Neural Network 3.1 Multi-task Learning (MTL) 3.2 Multi-task Learning in ASR Using DNNs and Our Formulas 3.3 Summary Chapter 4 Joint Acoustic Modeling of Phone and Grapheme: Information from Writing System 4.1 Graphemes as Acoustic Units in ASR 4.2 Joint Phone and Grapheme Acoustic Modeling by MTL-DNN 4.3 Analysis of Task Relationship

show abstract

Collecting and evaluating speech recognition corpora for 11 South African languages

Cited by 41 publications

References 18 publications

Using automatic alignment to analyze endangered language data: Testing the viability of untrained alignment

Using automatic alignment to analyze endangered language data: Testing the viability of untrained alignment

The Usefulness of Imperfect Speech Data for ASR Development in Low-Resource Languages

Multi-task learning deep neural networks for automatic speech recognition

Contact Info

Product

Resources

About