aiOla Research: Shaping the Future of Speech Intelligence

UmbraTTS is a new Text-to-Speech model that generates both speech and background audio together, creating more realistic, context-aware soundscapes. It uses a self-supervised approach to learn from unannotated recordings, overcoming the lack of paired training data. The result is high-quality, natural-sounding audio with fine control over environmental sound.

Target speaker extraction is extracting a specific speaker’s voice from a mixture of overlapping speech and background audio. In this work, we explore a simple yet effective approach to TSE using flow matching.

After setting new benchmarks in English, Spanish, French, and more, Jargonic V2 now leads in Japanese as well—delivering not just superior transcription accuracy, but also unmatched recall of specialized terms across industries like manufacturing, logistics, healthcare, and finance.

An enterprise-grade speech recognition model that outperforms all competitors across both academic benchmarks and real-world business environments. In comprehensive testing, Jargonic achieved the highest accuracy on standard datasets and superior jargon recognition capabilities, establishing it as the industry’s most accurate speech-to-text solution available. 

A novel multi-head efficient decoding approach for transformer-based Automatic Speech Recognition (ASR), improving inference speed and accuracy.

A privacy-focused speech recognition approach that enables entity recognition while anonymizing sensitive information, meeting enterprise-grade security and compliance requirements. 

An advanced framework that integrates named entity recognition (NER) into speech-to-text pipelines, enhancing real-time voice data processing.

An advanced adaptation model that enhances ASR performance in specialized domains by guiding recognition with contextual keyword injection.

A novel method for combining multiple language models to improve speech recognition across specialized industries, ensuring more accurate jargon recognition.

A cutting-edge technique enabling open-vocabulary keyword spotting using adaptive instance normalization to enhance real-time voice interaction and command execution.