Teaching keyword spotters to spot new keywords with limited examples

Awasthi, Abhijeet; Kilgour, Kevin; Rom, Hassan

doi:10.48550/arxiv.2106.02443

Cited by 1 publication

(5 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The user-defined KWS task can be viewed from two perspectives -as a classification task or a detection task. [13,14,19] solve user-defined keyword spotting task as a classification task. [19] classifies non-target keywords into multiple classes in pre-training, and re-trains the model on the target keywords with data augmentation.…”

Section: Related Workmentioning

confidence: 99%

“…[19] classifies non-target keywords into multiple classes in pre-training, and re-trains the model on the target keywords with data augmentation. [13] replaces the last linear layer with a randomly initialized linear layer during fine-tuning. [14] reinforces the model's representation capability by pre-training their model on the multilingual keyword dataset.…”

Section: Related Workmentioning

confidence: 99%

“…To overcome this shortage problem, few recent studies try to extract keyword data from a large-scale speech corpus using forced aligners [16]. For example, [13,15] and [14] extract keyword data from the LibriSpeech dataset [21] and the Common Voice dataset [22] respectively. [13,14,15] all use force-align transcripts to construct KWS datasets without verifying the aligned results.…”

Section: Related Workmentioning

confidence: 99%

“…For example, [13,15] and [14] extract keyword data from the LibriSpeech dataset [21] and the Common Voice dataset [22] respectively. [13,14,15] all use force-align transcripts to construct KWS datasets without verifying the aligned results. They also divide the datasets into training and test splits without considering the duplication of keywords between the sets.…”

Section: Related Workmentioning

confidence: 99%

“…However, the popular Google Speech Commands (GSC) dataset [12] only contains 35 classes, which is insufficient to facilitate good generalisation. To generate additional training data, previous works [13,14,15] extract keywords from Automatic Speech Recognition (ASR) datasets with a forced aligner [16] and then use them as training data. However, this approach does not guarantee the quality of the extracted keyword data.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Metric Learning for User-defined Keyword Spotting

Jung¹,

Youkyum²,

Park³

et al. 2022

Preprint

View full text Add to dashboard Cite

The goal of this work is to detect new spoken terms defined by users. While most previous works address Keyword Spotting (KWS) as a closed-set classification problem, this limits their transferability to unseen terms. The ability to define custom keywords has advantages in terms of user experience.In this paper, we propose a metric learning-based training strategy for user-defined keyword spotting. In particular, we make the following contributions: (1) we construct a large-scale keyword dataset with an existing speech corpus and propose a filtering method to remove data that degrade model training;(2) we propose a metric learning-based two-stage training strategy, and demonstrate that the proposed method improves the performance on the user-defined keyword spotting task by enriching their representations; (3) to facilitate the fair comparison in the user-defined KWS field, we propose unified evaluation protocol and metrics.Our proposed system does not require an incremental training on the user-defined keywords, and outperforms previous works by a significant margin on the Google Speech Commands dataset using the proposed as well as the existing metrics.

show abstract