The capability to convert spoken words into text is often underestimated, particularly with the rapid and accurate performance of the new AdaKWS model from aiOla, an Israeli tech startup founded in 2020 that specializes in speech recognition.
AdaKWS enhances OpenAI’s Whisper AI speech-to-text model, boosting keyword detection accuracy by 6.2% across 16 languages, and over 16% for English alone. Achieving a remarkable 94.6% accuracy in keyword spotting, it surpasses Whisper's 88.4% accuracy, according to metrics from aiOla. AdaKWS supports transcription in near real-time across 100 languages.
While these statistics might initially seem modest, they represent a significant leap from the 80th to the 90th percentile in accuracy. This upgrade transitions the technology from niche applications to broader use cases, even in highly regulated sectors such as healthcare and food safety.
Importantly, AdaKWS is also approximately 160 times faster at transcribing text than the Whisper-Large V2 model, according to aiOla's data.
“The ability to spot keywords enables automation of everyday processes across various industries, from filing parcel damage reports to completing safety inspections in food plants, transforming speech into action,” stated Amir Haramaty, CEO and co-founder of aiOla.
Diverse Enterprise Applications
While it's easy to associate speech-to-text AI with tasks like transcribing customer service calls, aiOla's technology is making strides in less conventional areas as well. In a media demonstration, Haramaty showcased the system's capability in a healthcare setting. A health tech speaker read off metrics from patient monitoring equipment, and the AdaKWS model automatically filled out a complex text form within seconds, eliminating the need for manual entry.
Additionally, aiOla has highlighted its application in monitoring supermarket refrigerator temperatures. By allowing human monitors to verbally report the readings, the system saves the client over 110,000 hours annually that otherwise would have been spent on manual data entry.
The potential for AdaKWS has garnered attention from industry leaders; Haramaty noted he received a call from Oracle CEO Larry Ellison, who expressed interest in applying the technology for healthcare records.
How AdaKWS Speech-to-Text Works
AdaKWS employs a cutting-edge keyword-spotting method that integrates effortlessly into business workflows, enabling automation via spoken commands. It operates as a machine learning algorithm that enhances existing speech-to-text models like OpenAI’s Whisper, interlacing itself between the model's encoder—responsible for interpreting spoken words—and the decoder, which converts audio into text.
“Our focus is optimization,” explained Joseph Keshet, aiOla's chief scientist.
Unlike conventional models that need extensive retraining for new keywords, AdaKWS swiftly adapts to accommodate over 100 languages and dialects. This adaptability makes it ideal for enterprise environments.
“Industry-specific terminology is prevalent and can dominate communication,” noted Haramaty. Keshet added, “Our system is trained to ensure accuracy for those keywords, represented within a latent space that effectively generalizes across languages."
AdaKWS is particularly beneficial for organizations where multilingual interactions occur, as it can be quickly tailored to an industry’s specific jargon. Users can submit keyword lists for the model to learn independently, detecting terms even without prior exposure to the spoken versions.
The model can be ready for use within hours, learning new languages, processes, and keywords rapidly.
A benchmark test across 16 languages demonstrated that AdaKWS not only exceeded Whisper's accuracy but also efficiently managed complex terms while using fewer computational resources. The underlying research was published in a scientific paper in September 2023.
Enhancing Business Operations
As businesses increasingly seek efficient and reliable solutions for managing complex data and communication tasks, aiOla’s AdaKWS represents a significant opportunity to streamline operations and reduce overhead. The technology is available through web and mobile applications, operating on a software-as-a-service (SaaS) subscription model based on user and use case.
aiOla's advancements in speech AI not only set a new industry benchmark but also pave the way for innovations that enhance AI integration into everyday business processes.
“I enjoy disruption, but I’ve come to realize that most people prefer not to be disrupted,” Haramaty concluded, emphasizing that AdaKWS aims to augment and improve existing business operations rather than replace them.
Ranking
