On a practical design of a low complexity speech recognition engine

Speech recognition for mobile devices

“…In Ortmanns et al (1997) the k-d tree and PSA techniques reduced the CPU-time required for likelihoods computations respectively down to 50.7% and to 39.9% compared to the baseline system without increase in WER. Vasilache et al (2004) present a phoneme based ASR, where HMM parameters μ and σ are represented by one out of 2 5 (for mean) and one out of 2 3 (for variance) quantizer levels. If in addition the input features will be quantized with a similar number of levels 2 5 , then the likelihoods can take only one out of 2 5 2 3 2 5 = 8192 fixed values, which can be precomputed and stored.…”

Rapid porting of ASR-systems to mobile devices

“…The exponent can be extracted by simple bit masking and shifting and already represents a coarse approximation of log 2 f . It can be turned into log b f to any base b by multiplication with log b 2 as shown in equation (5). The error is always lower than log b 2, but can be improved by additionally approximating the mantissa m. By using a polynomial of first (6) or second (7) degree the computation is reduced to a few multiplications and additions:…”

“…Especially the evaluation of the Gaussian Mixtures uses a lot of floating point operations for computing the Mahalanobis distances. The use of look-up tables can avoid these computations [5] either in part or completely. But on a PDA the necessary additional memory access is more expensive than doing the computation by means of integer arithmetic.…”

“…State-Clustered Tied-Mixture (SCTM) HMMs [4] can reduce the size of the acoustic model while accelerating its computation, but requires specifically trained models. For faster computation of Mahalanobis distances, Vasilache suggests to use look-up tables [5]. [6] uses integer computation for decoding and preprocessing instead.…”