**Version:**- 2.0, 25. Feb. 1992
**Hardware / OS requirements:**- Low Level SESAM Workstation
**Description:**-

SAM_SPEX provides an analytic basis for the definition of recogniser performance. The SAM project has adopted six speaker dependent speech parameter defined within the British Alvey project, MMI 132 STA. SAM_SPEX version 2.0 implements the extraction of these parameters from isolated words. The parameters measured are:- Speaking Rate: The duration of each token is calculated using an energy-based endpoint algorithm (time resolution 1msec) and related to the mean duration for all tokens of the same utterance.
- Energy: The ratio of the mean energy and the peak energy of the utterance.
- Pitch Frequency: The mean and variance of the pitch frequency estimated within the voiced parts of the endpointed utterance using a standard SIFT (Simple Inverse Filtering Tracking) algorithm.
- Voice Quality: The ratio of the energy between 0-2kHz and 0-fs/2Hz (fs = sampling frequency).
- Vocal Tract Area Distance: Calculated in two steps:
- The average area vector is computed.
- The vocal tract area distance is computed as the Euclidean distance between the average vocal tract area vector and the average of all vocal tract area vectors for tokens of the same utterance.

- Pattern Congruence: Calculated in two steps:
- The output values of an eighteen channels filter bank are computed every 10msec of a token.
- A dynamic time warping algorithm is used to calculate the distance between the token and all other tokens of the same utterance.

**Developing lab:**-

Jydsk Telefon

Sletvej 30

DK - 8310 Tranbjerg - J

DenmarkTel: (45) 86 29 33 66

Fax: (45) 86 29 90 68

**Contact:**- Sven Danielsen
**NB:**- This software is currently only available to SAM
partners as the definition of the parameters is under review and
an improved definition is in preparation.