Embedded industrial speech recognition for Modern Standard Arabic using HMM–GMM and phonetic discriminability analysis

Life Cycle Reliability and Safety Engineering · Springer Nature · 2026-11-19 · DOI: 10.1007/s10772-026-10299-0

Аннотация

Phonetic similarity between commands can be a source of recognition errors in small-vocabulary speech recognition systems. This study presents an embedded automatic speech recognition system for 35 Modern Standard Arabic industrial commands, combining a phoneme-based HMM–GMM recognizer with a phonetic vocabulary screening measure. The corpus contains 10,500 utterances from 30 speakers, with 2,100 utterances from 6 unseen speakers used for speaker-independent evaluation. The selected 3-state HMM with 8 Gaussian mixtures achieved 94.95% accuracy with 28.4 ms average decoding latency on a PC. On a Raspberry Pi 4, the same model preserved the 94.95% accuracy with 205.6 ms average latency, a real-time factor of 0.153, and 5.7 MB peak memory. Phonetic similarity was assessed using a Command Discriminability Index (CDI) based on normalized Levenshtein distance between phoneme sequences. Of the 595 command pairs, 134 low-CDI pairs (22.5%) accounted for 69.9% of the 103 pairwise recognition errors. A permutation test confirmed that this concentration was unlikely to occur by chance ($$p \approx {10^{ - 4}}$$). These results support CDI as a simple screening tool for identifying potentially confusable commands and demonstrate the feasibility of the HMM–GMM recognizer on resource-constrained embedded hardware.

Читать у издателя →

Похожие статьи