Multiple Time Scale Features in Multi-Stream ASR

Astrid Hagen

We show how multiple time scale information can be successfully combined using the "full combination" approach to multi-stream processing. To illustrate the procedure we take as input streams the commonly used raw, difference and second difference (plp) coefficients. Results show a significant performance improvement, in both clean speech and in real noise, compared to the usual approach of simply concatenating each stream into a single vector.