A Semi-spontaneous Dutch Speech Dataset for Speech Enhancement and Speech Recognition
A new 1.5-hour Dutch speech dataset recorded in noisy cafes challenges modern AI models.
A team from TU Delft has released DRES (Dutch Realistic Elicited Speech), a specialized dataset designed to stress-test AI speech systems under real-world conditions. The 1.5-hour collection features 80 Dutch speakers recorded in noisy indoor environments like cafes and public halls using a four-channel microphone array. Unlike clean studio recordings, DRES captures semi-spontaneous speech with background chatter and ambient noise, creating what the researchers call a "realistic elicited" scenario that better reflects how people actually use voice technology.
In their evaluation, the team ran eight state-of-the-art automatic speech recognition (ASR) models—including popular commercial and open-source systems—through the challenging DRES audio. Five of these models managed word error rates (WER) below 22%, demonstrating surprising resilience. However, when they applied five modern single-channel speech enhancement (SE) algorithms to clean up the audio first, they discovered something counterintuitive: none of the enhancement techniques improved ASR performance. In some cases, SE actually made recognition worse.
This finding directly challenges common assumptions in the speech processing field, where enhancement is typically seen as a necessary preprocessing step for noisy audio. The researchers argue their results emphasize the critical importance of testing AI models in realistic, multi-speaker environments rather than artificial lab conditions. The DRES dataset, submitted to Interspeech 2026, provides a new benchmark that could force developers to rethink how they train and evaluate speech AI for real-world deployment.
- DRES contains 1.5 hours of Dutch speech from 80 speakers recorded in noisy public spaces with a 4-channel microphone array
- 5 out of 8 tested ASR models achieved <22% word error rate despite challenging multi-speaker background noise
- Modern single-channel speech enhancement algorithms provided zero improvement to ASR performance, contradicting common practice
Why It Matters
Forces AI developers to test speech models in realistic noisy conditions, revealing significant gaps between lab performance and real-world usability.