Empirical evaluation of generative fusion decoding for Indonesian general and medical speech recognition

Abstract

This study empirically evaluates generative fusion decoding (GFD) with open-weight large language models for Indonesian automatic speech recognition (ASR) under two controlled offline conditions: general read speech (FLEURS-ID) and medical dictation (Harkit). The training-free byte-level fusion integrates three frozen language-model backends—SEA-LION, Sahabat-AI, and Mistral—with frozen Whisper-large-v3 during decoding. A development partition was used to select the static fusion weight and prompting condition, and a held-out test partition was used for final evaluation. On FLEURS-ID, the best GFD configuration reduced corpus-level word error rate (WER) from 5.91% to 4.17%, corresponding to a 29.36% relative reduction. On Harkit, it reduced WER from 15.89% to 12.30%, a 22.59% relative reduction. Paired utterance-level tests yielded p < 0.01 for both datasets. Several non-prompted configurations already improved over the unfused Whisper-large-v3 baseline, suggesting that the external language-model prior contributed more consistently than explicit prompting. Prompting produced limited changes on FLEURS-ID and larger but backend-dependent changes on Harkit. The results support GFD as a training-free decoding strategy under the evaluated conditions, while broader validation is required for noisy, spontaneous, streaming, and clinical deployment settings.