How NeuralSpace helps developers build language AI models in Oriya, German, and many more local languages
A few years ago, Ayushmann Dash and Felix Laumann realised that existing natural language processing (NLP) models performed poorly in locally spoken languages and the translations of the models weren’t really accurate.
Although NLP models existed in German (Felix’s mother tongue) and Oriya (Ayushman’s mother tongue) in 2018-2019, they did not come across models that were based on latest deep learning advancements.
So, the duo decided to build, a software as a service (SaaS) platform that offers developers a web interface and a suite of APIs for text and voice NLP tasks (language AI models) in more than 100 languages.
Local languages and improved accuracy in real-world use cases are the focus areas of the platform, which has received a $2.8-million Seed investment from Techstars, a Colorado-based global investment business.
“The Googles, Microsofts, IBMs and AWSs of this world offer models in these languages too, but they are often inaccurate in many real-world use cases. NeuralSpace is developing models with a local-languages-first approach. We make sure that our models perform well in locally spoken languages such as Gujarati, Telugu, and Gulf Arabic before we even train them on English, French, and Spanish,” explains Felix.
What does the platform offer?
The NeuralSpace platform consists of various language processing functionalities in more than 100 languages. It offers a suite of APIs for language AI models, with features such as entity recognition, language understanding, speech-to-text, text-to-speech, machine translation, and transliteration.
The functionalities come pre-trained with some basic understanding of every language and can be used in many industries. Developers can either consume the content directly or customise it to get their own unique AI model that accurately solves their business problems.
For example, in conversational AI solutions, such as chatbots and voicebots, the ‘language understanding’ offering helps recognise the customer’s intent and extracts relevant information from what is said or written to perform a relevant action. Natural language understanding can also predict the user’s sentiment, which can help analyse user behaviour. This helps create chatbots and voicebots that respond to the user in a more empathetic tone.
“For instance, if the user says, ‘I want to book a table at 8 pm tonight’, our language understanding would recognise that the customer wants to make a reservation (intent) and also the time (8 pm) and date (today),” explains Felix.
NeuralSpace can automatically detect the language a person writes in, translate between languages, transcribe what has been said into text, identify speakers in a meeting, and create a synthetic voice that reads out text.
The AI model that powers the platform is state-of-the-art and developed in-house along with some of the best researchers in this field, says Felix.
The technology is specially developed for locally spoken languages in India, Middle East, and Southeast Asia.
NeuralSpace is headquartered in London, UK and has offices in India and Tunisia. The startup’s team comprises 22 engineers, marketers and operators, who are all driven by the passion to break down the language barriers on the internet.
The founders met while they were students in Europe. Ayushman was studying computer science at the Technical University of Kaiserslautern in Germany, while Felix was at the Technical University of Denmark in Copenhagen.
“We greatly benefit from having 19 out of the 22 employees in India, natively speaking languages like Punjabi, Bengali, Oriya, Marathi, and Hindi. We also have an Arabic linguist working with us and are looking forward to hiring an even more global team,” says Felix.
The NLP startup recently joined Colarado-based Techstars, a global investment platform that provides access to capital and one-on-one mentorship, and received a $2.8-million Seed investment.
The company’s product is live today and is used by more than 300 companies. NeuralSpace’s key clients include Hello Ebbot, a conversational AI company for Scandinavian languages, and Eden AI, a marketplace to easily integrate and compare various machine learning vendors.
The startup offers a pay-as-you-go revenue model.
Market and competition
The global text-to-speech market was valued at $2 billion in 2020 and is believed to touch $5 billion by 2026, according to a MarketsandMarkets report.
NeuralSpace competes with Bengaluru-based Slang Lab, which develops augmented voice experiences for users. These experiences help users command their devices to perform actions without even having to touch the screen. NeuralSpace also competes with Deepdub, Neospaience, Papercup, and Lovo.
The startup believes its edge lies in the number of languages offered on its platform (more than 100 compared to 30 by competitors Google AI, Microsoft Azure Cognitive Services, and IBM Watson), the performance and accuracy of the models, and the fact that it offers an all-in-one platform with many language AI offerings.
“We have built a horizontally scalable platform that can automatically adapt to different usage patterns of our customers. During flash sales on Black Friday, our e-commerce models for language processing are heavily in use,” says Felix.
Currently, the NeuralSpace team is working on the development of all its core services, building customisation capabilities, and adding more languages, especially for automatic speech recognition (ASR) models.
“Many people speak a mix of two languages in their daily lives. For example, 'Hinglish’ (Hindi-English) or a mix of Arabic and English. We are heavily investing in building models that understand two languages together,” says Feliz.
In future, the startup wants to focus on building end-to-end solutions based on the individual services currently offered.
“For example, a voicebot is a combination of ASR, language understanding, and synthetic voice generation (also called ‘text-to-speech’). There are many more such examples and we want our customers to easily build all of them on NeuralSpace, in more than 100 languages, customised for their specific vocabulary,” says Felix.