Automatic Speech Recognition System Development in the "Wild"

Ragni, Anton; Gales, Mark J. F.

doi:10.21437/interspeech.2018-1085

Cited by 20 publications

(17 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For these results, no wideband adaptation was used, so the wideband data was downsampled to use the same features as the CTS data. Our results compare favourably to previously published results on the same dataset [21,23].…”

Section: Methodssupporting

confidence: 90%

Untranscribed Web Audio for Low Resource Speech Recognition

Carmantini¹,

Bell²,

Renals³

2019

Interspeech 2019

View full text Add to dashboard Cite

Speech recognition models are highly susceptible to mismatch in the acoustic and language domains between the training and the evaluation data. For low resource languages, it is difficult to obtain transcribed speech for target domains, while untranscribed data can be collected with minimal effort. Recently, a method applying lattice-free maximum mutual information (LF-MMI) to untranscribed data has been found to be effective for semi-supervised training. However, weaker initial models and domain mismatch can result in high deletion rates for the semi-supervised model. Therefore, we propose a method to force the base model to overgenerate possible transcriptions, relying on the ability of LF-MMI to deal with uncertainty. On data from the IARPA MATERIAL programme, our new semi-supervised method outperforms the standard semisupervised method, yielding significant gains when adapting for mismatched bandwidth and domain.

show abstract

Section: Methodssupporting

confidence: 90%

Untranscribed Web Audio for Low Resource Speech Recognition

Carmantini¹,

Bell²,

Renals³

2019

Interspeech 2019

View full text Add to dashboard Cite

show abstract

“…The original work was demonstrated on in-domain data-sets (conversational speech). Nevertheless, similar techniques have also been employed to use a seed model trained on out-of-domain data followed by SSL on in-domain data [1,17,18]. The latter is often obtained from data crawled from the web.…”

Section: Semi-supervised Training Using Lfmmimentioning

confidence: 99%

“…A total of 400 hours each for Lithuanian and Bulgarian are considered for the experiments. The results are reported on two sets: the dev set, which is part of the official Babel release, and the IARPA MATERIAL Analysis Pack 1 (Analysis) [17,22]. The dev set consists of only CS while the Analysis set contains the three domains: CS, NB and TB.…”

Section: Materials Data-set Setupmentioning

confidence: 99%

Incremental Semi-Supervised Learning for Multi-Genre Speech Recognition

Khonglah

Madikeri

Dey

et al. 2020

ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

In this work, we explore a data scheduling strategy for semisupervised learning (SSL) for acoustic modeling in automatic speech recognition. The conventional approach uses a seed model trained with supervised data to automatically recognize the entire set of unlabeled (auxiliary) data to generate new labels for subsequent acoustic model training. In this paper, we propose an approach in which the unlabelled set is divided into multiple equal-sized subsets. These subsets are processed in an incremental fashion: for each iteration a new subset is added to the data used for SSL, starting from only one subset in the first iteration. The acoustic model from the previous iteration becomes the seed model for the next one. This scheduling strategy is compared to the approach employing all unlabeled data in one-shot for training. Experiments using lattice-free maximum mutual information based acoustic model training on Fisher English gives 80% word error recovery rate. On the multi-genre evaluation sets on Lithuanian and Bulgarian relative improvements of up to 17.2% in word error rate are observed.

show abstract

“…Reference [12] introduces a crawler for YouTube to curate training dataset for ASR and demonstrates a 40% improvement in Word Error Rate (WER) on the Wall Street Journal test dataset. In [13], the authors address the problem of operating ASRs in a wide range of developing languages, such as Swahili, by proposing to automatically scrape audio from YouTube and Voice of America and use ASR system confidence scores as the primary metric for the model components. The creators of the VoxCeleb1 and VoxCeleb2 datasets [14], [15], crawled YouTube to construct the datasets, which are now widely used in the field of Speaker Recognition [16].…”

Section: Related Workmentioning

confidence: 99%

Automated Techniques for Creating Speech Corpora from Public Data Sources for ML Training

Drabeck¹,

Ramanan²,

Woo³

et al. 2020

IJMLC

View full text Add to dashboard Cite

For machine learning (ML) to work well, there is a need for large amounts of good quality training data. Obtaining such data is often the key bottleneck for the entire ML development process. Using humans to do explicit collection has been the main approach, but this tends to be expensive and time-consuming. Therefore, there is significant interest in creating alternative data collection techniques. We explore these alternative data collection techniques in the context of speech data in this paper. We were initially motivated by the problem of wake word engine training, where we need a large number of utterances for specific wake words. Given that there are already large public repositories of media data (e.g., YouTube, DailyMotion), we were curious as to how feasible it is to find the utterances that we need. Our results are encouraging as we found many different types of words can readily be found and downloaded in the quantity and quality needed to create training corpora for DL training. Usually > 30% of the found words are suitable for corpus creation. Greater than 80% of the top 10,000 ranks words and > 50% of the top 20,000 words we selected easily produced > 5000 found words, which is sufficient to train a high quality Wake Word Engine. Besides general words, we specifically looked for words used in wake word engine construction such as Name/Place/Product Name. Here, again, we find most common names/places/products return more than a sufficient number of words for corpus creation. Only uncommon names and places (like Atticus or Maximus) are difficult to find in sufficient quantities for corpus creation. We demonstrate a wake word engine trained from words we found in YouTube has the equivalent performance to one trained with traditional human collected words. Even though we were focused on wake words, our approach is general. It can be applied to create speech corpus for various purposes.

show abstract

Automatic Speech Recognition System Development in the "Wild"

Cited by 20 publications

References 22 publications

Untranscribed Web Audio for Low Resource Speech Recognition

Untranscribed Web Audio for Low Resource Speech Recognition

Incremental Semi-Supervised Learning for Multi-Genre Speech Recognition

Automated Techniques for Creating Speech Corpora from Public Data Sources for ML Training

Contact Info

Product

Resources

About