Creating a task-specific speech recognition dataset is essential for developing speech recognition applications in low-resource languages. Such applications have uses in agriculture, finance, healthcare, and others, and benefit individuals with low literacy. However, a significant challenge is the high cost of data creation. While there is some work around cost-effective dataset selection, there is little to no work on building a cost-effective dataset for a task from scratch. Our work contributes to the latter. We created a speech recognition dataset from scratch and conducted two major sets of experiments. The first aimed to observe the effect of different datasets of the same size on model performance. Our results confirmed that the same amount spent collecting data can have vastly different results. The second experiment analyzed the effect of token sequence overlap between target and training data since a natural and intuitive approach to building a dataset from scratch for task would be having the task tokens occur in the training data. Our experiments showed that token sequence overlap was not the primary factor influencing model performance. Our work provides a counter-intuitive insight into building speech recognition datasets from scratch in low-resource settings and shows the need for further investigation.

Grammatical error correction (GEC) aims to improve text quality and readability. Previous work on the task focused primarily on high-resource languages, while low-resource languages lack robust tools. To address this shortcoming, we present a study on GEC for Zarma, a language spoken by over five million people in West Africa. We compare three approaches: rule-based methods, machine translation (MT) models, and large language models (LLMs). We evaluated GEC models using a dataset of more than 250,000 examples, including synthetic and human-annotated data. Our results showed that the MT-based approach using M2M100 outperforms others, with a detection rate of 95.82% and a suggestion accuracy of 78.90% in automatic evaluations (AE) and an average score of 3.0 out of 5.0 in manual evaluation (ME) from native speakers for grammar and logical corrections. The rule-based method was effective for spelling errors but failed on complex context-level errors. LLMs—Gemma 2b and MT5-small—showed moderate performance. Our work supports use of MT models to enhance GEC in low-resource settings, and we validated these results with Bambara, another West African language.

In this paper, we present our system for assigning an emoji to a tweet based on the text. Each tweet was originally posted with an emoji which the task providers removed. Our task was to decide out of 20 emojis, which originally came with the tweet. Two datasets were provided - one in English and the other in Spanish. We treated the task as a standard classification task with the emojis as our classes and the tweets as our documents. Our best performing system used a Bag of Words model with a Linear Support Vector Machine as its’ classifier. We achieved a macro F1 score of 32.73% for the English data and 17.98% for the Spanish data.