Interspeech'2005 - Eurospeech

Lisbon, Portugal
September 4-8, 2005

Comparison of Low Footprint Acoustic Modeling Techniques for Embedded ASR Systems

Jussi Leppänen, Imre Kiss

Nokia Research Center, Finland

In this paper we compare the performance of speech recognition systems based on hidden Markov models (HMM) with quantized parameters (qHMMs) and subspace distribution clustering hidden Markov models (SDCHMMs). Both of these HMM types provide similar performance as continuous density HMMs, but with significantly reduced memory requirements (approximately 90% less memory was needed to store the HMM densities). The experiments show that on a small vocabulary isolated word recognition task, SDCHMMs outperform qHMMs in clean conditions. However, when noisy test data is used and adaptation is enabled qHMMs outperform the SDCHMMs. In addition, the experiments show that as low as 3-bit feature quantization can be used with both qHMMs and SDCHMMs without sacrificing recognition performance.

Full Paper

Bibliographic reference.  Leppänen, Jussi / Kiss, Imre (2005): "Comparison of low footprint acoustic modeling techniques for embedded ASR systems", In INTERSPEECH-2005, 2965-2968.