How NeuralSpace is driving low-resource language processing

London-based NeuralSpace is offering a no-code platform to help enterprises train and deploy deep learning models capable of processing low-resource languages.

For years, natural language processing (NLP) has been enabling AI-driven systems to understand and produce text/speech in the same way humans do. The technology is heavily used in areas such as speech recognition, but its demand has particularly grown in light of the pandemic and the accelerated speed of digital transformation. A recent survey from John Snow Labs and Gradient Flow found that 60% of tech leaders saw their NLP budgets increase by at least 10% compared to 2020, while 33% witnessed over 30% increase.

But, here’s the thing, training and deploying NLP models is a challenge in itself. Organizations need significant capital and compute resources to train language models, and if they manage to arrange it, there’s a good chance they could run into data hurdles. After all, one cannot find as much data and research for low-resource languages, predominantly spoken in Asia, the Middle East and Africa, as they could for high-resource ones such as English, Spanish, and French.

NeuralSpace’s no-code platform

To bridge this gap, NeuralSpace has developed a no-code modular toolkit that offers a suite of APIs to help enterprises train and deploy AI models for language processing applications. It supports more than 80 languages, including low-resource ones, and leverages modeling techniques such as transfer learning, multilingual learning and data augmentation to work with only a handful of text or speech data.

“The platform can be accessed from the CLI (command-line interface) for developers, by REST APIs for an easy web or online product integration and even by a no-code GUI so that training, testing, and deploying a model in production can be done without writing one line of code,” the company said in a statement.

In all, NeuralSpace supports eight processing functions – conversational intelligence, machine translation, entity recognition, transliteration, data generation, language detection and speech-to-text as well as text-to-speech conversion. It leverages a language-agnostic automated NLP approach that automatically figures out which pipeline, features and model parameters will produce the best results for a given dataset. 

And when the work is complete, an automated MLops module scales the deployed models for higher availability and throughput. 

“Currently, enterprises need to combine multiple tools to automate language-specific tasks, such as answering frequent customer queries. With NeuralSpace, everything is under one hood for almost any language in the world. It is a one-stop solution for language technology, both in voice and text formats,” Felix Laumann, the CEO of NeuralSpace, told Venturebeat.

“The Artificial Intelligence (AI) models that power our platform are state-of-the-art and developed in-house together with some of the best researchers in this field. They are ready for developers to either customize or consume out of the box with a few clicks,” he added.

Competition

Another player in the same space is Hugging Face, a company that provides open-source NLP technologies. Felix claims that the number of NLP models and datasets on Hugging Face is huge, but only a small portion (5%) of that is available for low-resource languages. Plus, in order to use those offerings, one needs to have sufficient knowledge of NLP.

“In India, Reverie Language Technologies and Skit AI are competitors, but they mainly focus on Machine Translation and Speech Recognition, respectively. NeuralSpace offers the first platform that combines multiple of these functionalities in more than 80 languages,” he added.

Funding

Initially, NeuralSpace was just a side project, but in 2019, it was formally transformed into a company. Today, the startup announced it has raised $1.7 million in a seed round, led by Merus Capital. 

With this funding, NeuralSpace plans to improve its no-code NLP platform and add state-of-the-art speech models for the same 80 plus languages. It also plans to bring the ability to trade domain-specific pre-trained NLP models on a marketplace. 

“We aspire to establish ourselves as the go-to provider for NLP services in low-resource languages. The notion will always be on how software and mobile application developers can implement NLP features into their products with as little knowledge as possible about ML and Data Science,” Felix said.

“We plan to work with the best researchers and UX designers in this field to provide an unmet experience to developers that allows them to integrate a voice command system, an automated summarization, or a sentiment analysis, besides many other language technologies, just with a click of a button,” he added.