Recent advancements in large-scale pre-trained automatic speech recognition (ASR) foundation models (e.g., Whisper) have exhibited remarkable performance in speech processing tasks. But fine-tuning such models for low-resource languages can be computationally expensive and prone to overfitting. Prompting methods offer a solution by designing specific prompts in the inputs that guide the model’s behavior for targeted tasks, facilitating parameter-efficient adaptation. This paper presents the first exploration of various prompt tuning methods and optimized strategies for low-resource ASR based on Whisper. Moreover, we propose a shallow integration method to utilize the advantage of deep prompt tuning and reparametrization. Extensive experiments on the Common Voice and FLEURS datasets show the competitive performance of prompt tuning compared to full fine-tuning and Lora with fewer trainable parameters. Notably, the shallow integration strategy yields impressive results, especially for small models.