Massive multiple-input multiple-output (mMIMO) regime reaps the benefits of spatial diversity and multiplexing gains, subject to precise channel state information (CSI) acquisition. In the current communication architecture, the downlink CSI is estimated by the user equipment (UE) via dedicated pilots and then fed back to the gNodeB (gNB). The feedback information is compressed with the goal of reducing over-the-air overhead. This compression increases the inaccuracy of acquired CSI, thus degrading the overall spectral efficiency. This paper proposes a computationally inexpensive machine learning (ML)-based CSI feedback algorithm, which exploits twin channel predictors. The proposed approach can work for both time-division duplex (TDD) and frequency-division duplex (FDD) systems, and it allows to reduce feedback overhead and improves the acquired CSI accuracy. To observe real benefits, we demonstrate the performance of the proposed approach using the empirical data recorded at the Nokia campus in Stuttgart, Germany. Numerical results show the effectiveness of the proposed approach in terms of reducing overhead, minimizing quantization errors, increasing spectral efficiency, cosine similarity, and precoding gain compared to the traditional CSI feedback mechanism.