Lately, we have been developing our own Finnish speech-to-text model and now it’s ready! We’ve designed the Finnish speech-to-text model to help large companies detect business-critical patterns from existing phone calls, such as sales leads, quality leaks, unhappy and leaving customers. Analysing phone calls helps sales managers to find new sales leads (upsell and win-back leads) from existing conversations in near real-time while also providing customer service managers with a tool to help improve customer service quality.
Converting customer phone calls into text is a necessary step to find business-critical patterns. After that, all calls become searchable and AI-based models automatically detect each call’s sentiment, topic, etc. This step spawns actionability which could mean contacting clients, improving services, taking advantage of sales opportunities, etc. You can see the Feelingstream conversation analytic tool’s value creation process in the following graph:
There are a couple of reasons why we developed our own speech to text model instead of using cloud-based services:
- Higher accuracy. Our speech to text (speech-to-text or ASR – Automatic Speech Recognition) model is more accurate because we trained it for customer conversations in Finnish in specific sectors. We kept our market in mind while collecting training data for the model.
- Adaptability. Our speech-to-text model is easily adaptable to customer-specific words. For example, if the model doesn’t recognize specific product names, we could easily fine-tune it to increase accuracy.
- On-premises. Our customers can use our speech-to-text model on-premises without any external connections.
Adopting the Finnish speech-to-text model
We built the model in the Finnish language based on data collected from public sources. The crowdsourcing is great to run this project. First tests on real customer phone calls have been very positive. Even if we do not adapt the model to specific customer’s phone calls, it has a lower word error rate than Google’s Speech to Text API. Feelingstream’s FIN ASR model has the added advantage that we can decrease its error rate even further.
If you would like to hear more about detecting business-critical patterns in customer conversations, contact us or come see us at AI Monday in Helsinki on 21st October 2019. Please see the recording in English (the article is in Finnish, but the video is in English).
Risto Hinno is a Lead Feeling Scientist at Feelingstream and takes care of developing AI-based models for detecting business-critical patterns from customer conversations.
