Can an initial phoneme-based algorithm improve automatic naming latency detection during picture naming tasks? A feasibility study

• Speech envelope of audio signal fosters initial phoneme-specific word onset detection. • Speech onset detection by Kaldi speech recognition toolkit can be used as reference time in naming latency detection. • Second and third phonemes might influence speech envelopes. • An initial phoneme-based optimization approach outperforms existing speech onset detection tools. Picture naming tasks are widely used to evaluate language processing in aphasia. Automatically detected naming latencies (aNL) often differ from the gold standard, manually detected latencies (mNL). Initial phonemes (InitP) influence aNL detections. The study aimed to develop and explore an InitP-based optimization approach to enhance automatic naming latency detection. An algorithm was developed to optimize several parameters of the word speech envelope for 16 InitPs using German training data from 134 healthy subjects and 31 persons with aphasia. These parameters were applied to test data to assess InitP influence. The resulting naming latencies (FHNW_NL), detected by the newly developed algorithm, were compared to manually detected latencies (mNL) and to speech onsets identified by Kaldi and Chronset. The mean signed difference between FHNW_NL and mNL falls within the ± 15 ms inter‑rater variability interval. Differences from Kaldi and Chronset to mNL were over twice and four times higher, respectively, than FHNW_NL. Persons with aphasia showed higher variances than healthy subjects. This study shows that InitP-based optimization of naming latency detection parameters improves aNL detection accuracy. This method could also benefit other research and clinical fields.