Objective. The brain–computer interface is currently experiencing a surge in the number of recording channels, resulting in a vast amount of raw data. It has become crucial to reliably detect neural spikes from a large population of neurons in the presence of noise, in order to constrain the transmission bandwidth. Approach. We investigate various time–frequency analysis methods for spike detection, followed by an exploration of energy operators amplifying spikes and signal statistics for adaptive thresholding. Subsequently, we introduce a precise and computationally efficient spike detection module, leveraging stationary wavelet transform (SWT), Teager energy operator, and root-mean-square calculator. This module is capable of autonomously adapting to different levels of noise. The SWT effectively eliminates high-frequency noise, enhancing the performance of the energy operators. The hardware computational process is simplified through the use of the lifting scheme and a channel-interleaving architecture. Main results. We evaluate the proposed spike detector with adaptive threshold on the publicly available WaveClus datasets. The detector achieves an average accuracy of 98.84%. The application-specific integrated circuit (ASIC) implementation results of the spike detector demonstrate an optimized interleaving channel of 8. In a 65 nm technology, the 8-channel spike detector consumes a power of 0.532 μW Ch−1 and occupies an area of 0.00645 mm2 Ch−1, operating at a 1.2 V supply voltage. Significance. The proposed spike detection processor offers one of the highest accuracies among state-of-the-art spike detection methods. Importantly, the ASIC explores the considerations in the scalability and hardware costs. The proposed design provides a systematic solution on spike detection with adaptive thresholding, offering a high accuracy while maintaining low power and area consumptions.