Data Augmentation for Low-Resource Quechua ASR Improvement

Abstract

Automatic Speech Recognition (ASR) is a key element in new services that helps users to interact with an automated system. Deep learning methods have made it possible to deploy systems with word error rates below 5% for ASR of English. However, the use of these methods is only available for languages with hundreds or thousands of hours of audio and their corresponding transcriptions. For the so-called low-resource languages to speed up the availability of resources that can improve the performance of their ASR systems, methods of creating new resources on the basis of existing ones are being investigated. In this paper we describe our data augmentation approach to improve the results of ASR models for low-resource and agglutinative languages. We carry out experiments developing an ASR for Quechua using the wav2letter++ model. We reduced WER by 8.73% through our approach to the base model. The resulting ASR model obtained 22.75% WER and was trained with 99 hours of original resources and 99 hours of synthetic data obtained with a combination of text augmentation and synthetic speech generation.

G-Augment: Searching for the Meta-Structure of Data Augmentation Policies for ASR

“…Data augmentation using either labelled or unlabelled data [1,2] has been used to alleviate this data scarcity problem. One promising approach is speech data synthesis, which recently contributed to significant progress in domain adaptation [3], medication names recognition [4], accurate numeric sequences transcription [5], low-resource languages [6], etc. Most research in this areas has focused on text-tospeech (TTS) generation [7,8,9,10] and audio augmentation [11,12,13], and their effects on the resulting ASR accuracy.…”

“…Zevallos et. al. shows a good improvement from the non-augmented dataset [6]. In [6], synthetic utterances were created by first replac-*Equal contribution.…”

“…Our neural model based TA methods utilize BART, a large pre-trained model [15] which has proven useful in large scale synthetic text generation with minimal human effort [16]. We utilize mask filling to create synthetic utterances where it can find similar words from the large text corpus it was pre-trained on, unlike [6] which finds words only from the smaller seed data training corpus. These TA methods are compared on the open domain LibriSpeech, a Photo Search domain with short, structured utterances, and a Question and Answering (Q&A) domain with relatively freestyle dictations.…”