Headlines keep popping up about problems with the speech recognition of AI solutions. But how big are the complications surrounding dialects and artificial intelligence really? And how can a company ensure that the technology is reliable despite linguistic diversity?
The obstacles in speech recognition
Just recently, a case involving a German company made the headlines. The automated call support failed due to the caller’s Franconian dialect, who ultimately had to clarify his request in person at the company’s branch. This problem is not an isolated incident – different dialects and accents are still difficult for most software solutions to understand.
The steps behind
Back to the beginning: speech recognition is a technology for converting spoken language into text. Algorithms and training processes are used to train these models with a large amount of speech data. The quality of this process is crucial for the accuracy of the solution. Standard high-level languages are often used, which makes it difficult to recognize some regional dialects and accents.
Quality in training
For us, the key factor in reliable and functional speech recognition is precise training. At LinkThat ECCO, we have placed particular emphasis on using a wide variety of speech data – with success:
Our artificial intelligence for telephony and voice channels has been successfully in use at the Austrian Health Insurance Fund (ÖGK) for three years. The ECCO Attendant supports callers in the queue and takes appropriate action. Despite the linguistic diversity of the federal states, our technology works flawlessly with a few exceptions. As a result, 87% of callers to ÖGK are routed directly to the right destination (the rest are placed in the standard queue for manual transfer). The time saved is therefore enormous thanks to the accuracy.
Keyword Spotting in real-time
ECCO also pays particular attention to keyword spotting, which provides real-time support during calls. Important keywords are defined in advance, which then trigger helpful actions during the call. If specific words or phrases are used, ECCO becomes active immediately.
We opted for this approach of keyword recognition, among other things, because it also works very well in a country like Austria with its many facets and dialects. Even without special training for the AI, predefined words are recognized extremely well. The simple reason: if callers have a question about a specific issue, they usually make an effort to pronounce it clearly.
Let’s stay with the ÖGK example: an insured person wants to find out more about breast cancer screening. Even in Austria, it is unlikely that this key word will be spoken in a strong dialect. By focusing on keywords, you automatically achieve good success rates. Special topics can of course still be “learned”.
A technology with great potential
We are constantly working on improving and expanding our products. For example, ECCO will soon be updated to include a live transcription with a call summary.
The development of precise speech recognition is a challenge that we face with passion and commitment. With LinkThat ECCO, we support companies in communicating more efficiently and effectively with their customers.