Research · Speech-to-Text
This research presents the design and development of a neural Speech-to-Text system for Krio, a widely spoken language in Sierra Leone, as the first phase of a broader initiative to develop speech technologies for African languages.
Abstract
Automatic Speech Recognition, commonly known as Speech-to-Text, enables computers to convert spoken language into written text. Although recent advances in deep learning have significantly improved ASR for high-resource languages, most African languages remain underserved due to limited speech datasets and linguistic resources. This research presents the design and development of a neural Speech-to-Text system for Krio and outlines a scalable path toward expanding ASR to other African languages.
Speech recognition has become an essential component of modern artificial intelligence, enabling applications such as virtual assistants, automatic transcription, call-center automation, real-time captioning, and multilingual communication. However, these technologies are primarily available for high-resource languages, leaving many African languages without reliable speech recognition systems.
Developing high-quality Speech-to-Text systems for Krio presents several challenges, including limited publicly available speech datasets, insufficient diversity in existing recordings, variations in accents, pronunciation, and speaking styles, background noise in real-world recordings, lack of standardized benchmarks for evaluation, and limited pretrained models optimized for African languages.
The objectives of this research are to develop a large, high-quality Krio speech corpus, build a robust neural Speech-to-Text model, design an automated speech preprocessing pipeline, improve transcription accuracy across different speakers and recording conditions, and create a scalable framework for extending ASR to additional African languages.
Recent progress in ASR has been driven by deep learning architectures such as Connectionist Temporal Classification, Transformer-based speech models, encoder-decoder architectures, self-supervised speech representation learning, and multilingual speech foundation models. State-of-the-art systems such as Whisper, MMS, and Wav2Vec 2.0 demonstrate strong performance on multilingual speech recognition, but adapting them to low-resource African languages requires high-quality datasets and language-specific fine-tuning.
A Krio speech dataset was developed using recordings from native speakers spanning male and female voices, different age groups, multiple regional accents, conversational speech, public speaking, read speech, number reading, everyday expressions, and questions and commands. Each audio recording was paired with an accurately verified transcript to create supervised training data.
A comprehensive preprocessing pipeline was developed to improve training quality through audio format standardization, resampling to a consistent sampling rate, voice activity detection, silence trimming, segmentation of long recordings into shorter utterances, transcript normalization, removal of low-quality recordings, and dataset validation.
The research explored modern neural ASR architectures suitable for low-resource languages, including Whisper, MMS-ASR, Wav2Vec 2.0, and Conformer-based ASR models. Transfer learning was used to adapt pretrained multilingual speech models to Krio, reducing the amount of labeled data required while improving recognition performance.
The model development process included audio preprocessing, transcript normalization, feature extraction, tokenization, fine-tuning pretrained models, validation on held-out data, hyperparameter optimization, and model checkpointing. Training emphasized speaker diversity and varied recording environments to improve generalization.
Several challenges were encountered during development, including data scarcity, speaker diversity, recording quality issues such as background noise and inconsistent microphones, long audio recordings that required segmentation, and language variation involving regional pronunciation differences, borrowed vocabulary, and code-switching with English.
The Speech-to-Text system was evaluated with objective metrics such as Word Error Rate, Character Error Rate, Sentence Error Rate, inference latency, and Real-Time Factor, alongside qualitative review of transcription accuracy, handling of different accents, recognition under noisy conditions, performance on conversational speech, and general usability.
The developed Speech-to-Text system supports numerous real-world applications, including automatic meeting transcription, voice typing, AI-powered customer support, real-time captioning, broadcast and media transcription, digital government services, healthcare documentation, educational platforms, and accessibility tools for individuals with hearing impairments.
Future research will focus on real-time streaming Speech-to-Text, speaker diarization, noise-robust recognition, code-switching recognition between Krio and English, multilingual ASR models, end-to-end speech translation, and expansion to Mende, Temne, Limba, Fula, and other African languages. This research demonstrates that robust Speech-to-Text systems can be developed for low-resource languages through high-quality dataset creation, careful preprocessing, transfer learning, and modern neural architectures.