- IMD Business School
News Stories · Entrepreneurship

Spitch’s robo-bankers are learning Swiss German!

For the fourth edition of the Fintech Chain Mail, we met with Alexey Popov – CEO at SPITCH to discuss Natural Language Processing, a form of artificial intelligence that extracts data and insight from speech.
November 2020
How does Spitch serve clients with Natural Language Processing (NLP) technologies?

There are two ways Spitch serves its customers, which range from large banks with millions of customers to smaller cantonal banks, with out-of-the-box, high-quality and affordable offerings.

Firstly, our Voice Assistants provide real-time speech recognition solutions. We deliver mainly out-of-the-box ‘Open Dialogue’ products based on artificial intelligence and neural networks. Our systems are built to understand the intent of the caller and the context behind their speech with all its’ features (as opposed to only transcribing and interpreting text). By contrast, legacy speech-to-text technologies only transcribe words, which has proved ineffective. Such old technologies would fail and immediately lose the callers’ intent if, for instance, words are pronounced differently to keywords the system was trained on. Solutions run in an omni-channel mode, so people can start conversations verbally and end up in a chat or vice versa.

Secondly, we offer Speech Analytics, which run in real-time or in batch mode. Companies have huge amounts of voice data, but they generally do not use this treasure to improve their business, enhance customer experience and retention. Our technology makes it possible to transition from manual processes to comprehensive, automated analytics which expose important findings and dependencies. We can perform sentiment analysis and pick up on differences between genuine and sarcastic ‘thank you’, for instance, which could help companies measure and track customer satisfaction. We can also analyze variations in tone over a conversation, such as changes in customers’ sentiment from negative to positive and conclude on the whether the conversation was positive or negative overall. For banking clients for instance, we also carry out communications surveillance of traders’ and agents’ compliance to scripts, using real-time speech analytics. This helps banks and brokerages better serve customers and supports compliance and forensic teams in their functions.

How does Spitch differentiate itself among established players in NLP like Amazon, Apple and Google?

We do not try to offer every solution possible to everyone – we only serve clients whose businesses we fully understand or where we can rely on trusted partners with very short time to market. These partners can adopt our solutions to create the same out-of-the-box solutions for their customers. It is unlikely that big cloud-based international platforms can approach big bank with a full understanding of their businesses and the local dialects they operate in.

There are also trust and regulatory issues that might influence the willingness or ability of banks, insurance or credit card companies to divulge customer data to big technology companies. This is especially the case where these technology companies could become banks’ future competitors. As such, banks’ customer data is a key competitive advantage which they are unlikely to surrender. Spitch delivers solutions on-premise or on clients’ private clouds so clients can keep and process customer data on their terms.

While Spitch pursues big name clients, we recognize they could be conservative due to bad experiences with legacy solutions. On the other hand, the majority of NLP technologies are too expensive for many small businesses due to high adaptation and integration costs. Our strength lies in our ability to deliver products which can run right out-of-the-box, without requiring expensive bespoke development. This puts Spitch at an advantage since our products are more affordable than extensive technology projects, allowing us to properly serve small and agile companies.

From these successes, we expect to draw more big fish. We have just completed a fundraising round having attracted new investors, and this should enable us to scale. Also, Spitch is very proud to be mentioned in the Gartner Market Guide for text-to-speech technologies among about 16 other companies including Amazon Web Services, Google, IBM and Microsoft. In addition, Gartner named Spitch a “Cool vendor” in specially devoted report.

How does Spitch train its algorithms – what is the source of data that goes into the system and how manual is the process?

Swisscom was our first partner in Switzerland, where we were able to process call centre conversations. From this procured data, we generated significant language knowledge, including in French and Swiss German. Following this experience, we continue to develop our database, models and systems using new customers’ and artificially collected data to continuously enhance our solution offerings for customers over time, while respecting all data privacy issues and regulations.

We have also created an eco-system of Swiss locals to work with us on a temporary basis as annotators. They can work fully remotely, which is key in these times, using our frameworks and tools to further process the data within the system and improve its learning. When analysing data on customer interactions for instance, our system educates itself based on historic conversations with the help of our analysts.

Some NLP trials, such as Microsoft’s abusive chatbot from 2016, have yielded bad results – how do you see the limitations of properly training algorithms without human intervention being solved?

We never let our chatbots run on their own, as that is dangerous indeed. Our process with clients begins with intent recognition and automating very clear frequent customer queries, as this step introduces limited integration and disruption. Next, we very carefully switch on the integration into clients’ back-end system to handle more comprehensive queries.

This chatbot or automation solution assists our clients’ agents and ultimately handles some cases solely. However, generally we design teams comprising humans and chatbots working together. We are very careful in training our system – which can be thought of as a puppy. If it is not well trained and looked after, it could become a wolf rather than a dog.

How does adoption of NLP in Switzerland compare to other countries?

The roll-out of NLP has been slow in Switzerland, unlike other countries such as the U.S. where adoption was heavier and occurred earlier. Unfortunately, in some of these countries, negative customer experience with ineffective chatbots and other solutions based on old technologies has hugely eroded customer perceptions. In Switzerland, people’s first experience is likely to be more positive given advances that have been made with NLP technologies.

In Switzerland, Spitch is the leading NLP player, because we operate in major European languages (English, French, Italian, and high German), as well as the Swiss German dialects which are very important locally. As we roll out Spitch Solutions, we learn from the mistakes of others to improve our solutions.

Interview conducted by Stephanie Hurry, Conti Matteo, Olabisi Ayodeji, Emon Goswami