8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003


Low Memory Acoustic Models for HMM Based Speech Recognition

Tommi Lahti, Olli Viikki, Marcel Vasilache

Nokia Research Center, Finland

In this paper, we propose a new approach to reduce the memory footprint of HMM based ASR systems. The proposed method involves three steps. Starting from the continuous density HMMs, mixture variances are tied using k-means based vector quantization. Next, the re-estimation of the resulted models is performed with tied variances. Finally, scalar quantization is performed for the mean components of the models. With the proposed method, a memory saving of 77.6% was achieved compared with the original continuous density HMMs and 23.0% compared to the quantized parameter HMMs, respectively. The recognition performance of the resulted models was similar to what was obtained with the original continuous density HMMs in all tested environments.

Full Paper

Bibliographic reference.  Lahti, Tommi / Viikki, Olli / Vasilache, Marcel (2003): "Low memory acoustic models for HMM based speech recognition", In EUROSPEECH-2003, 2489-2492.