Automatic Language Discrimination.
Abstract
A language discrimination algorithm was developed that utilizes single steady-state phoneme-like reference sounds. The algorithm makes use of the relative frequency of occurrence of these reference sounds and nearest-mean decision strategy. The reference sounds were determined interactively and have been shown to have good language specificity. The number of reference sounds required is small enough to allow real-time operation. Standardization of each speaker's long-term average spectrum removed much of the data variation that was due to variations in data recording conditions. This standardization provided a 41 percent decrease in the number of test speaker misclassifications. In a performance test involving 50 test speakers of five languages, 80 percent correct classification was achieved. Excellent discrimination among L1, L3, and L5 was attained, while additional references with more language specificity are needed for discriminating between L2 and L4. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1978
- Accession Number
- ADA050841
Entities
People
- George R. Doddington
- R. Gary Leonard
Organizations
- Texas Instruments