Optimizing Whisper models for Turkish automatic speech recognition: A comprehensive study on parameter-efficient fine-tuning and model compression

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Turkish, as a morphologically rich and agglutinative language, poses distinctive challenges for automatic speech recognition (ASR) systems. Although large foundation models such as OpenAI’s Whisper exhibit strong multilingual performance, their out-of-the-box accuracy on Turkish remains limited and requires systematic adaptation. In this study, we investigate three complementary fine-tuning strategies for adapting Whisper-Medium and Whisper-Large-v3 models to Turkish speech, with the aim of characterizing the trade-offs between efficiency, accuracy, and specialization. Specifically, we evaluate (i) parameter-efficient LoRA applied to attention projections, (ii) extended LoRA covering all linear layers, and (iii) decoder-focused specialization with encoder freezing, combined with aggressive model compression via 4-bit NF4 quantization and 10% L1 pruning. These strategies collectively form a principled ablation suite that enables a structured exploration of adaptation depth for a morphologically complex language where full fine-tuning is often prohibitively expensive. The models are trained on the Turkish Speech Corpus (ISSAI_TSC_218), comprising 218.2 h of speech and 186,171 utterances, and evaluated on a held-out test partition. Post-training optimization using CTranslate2 allows efficient deployment with FP16 and INT8 quantization. Experimental results show that the decoder-focused INT8 configuration achieves the best performance, yielding a word error rate of 14.89% and a character error rate of 3.96%, while maintaining real-time inference capability (RTF = 0.035 on GPU and RTF = 0.6–0.8 on CPU). This corresponds to a relative WER reduction of 20.8% compared to the base Whisper-Large-v3 model. By enabling training of its most compressed configuration (4-bit decoder-focused fine-tuning) on consumer-grade hardware (4 GB VRAM) and deployment on standard CPUs, the proposed framework provides practical guidelines for resource-efficient Turkish ASR. Cross-corpus evaluation on an independent Turkish speech dataset shows that these gains generalize beyond the training corpus, though validation on noisy, spontaneous, or accented speech remains an open direction before deployment in acoustically challenging, real-world settings.