Introduction Electrocardiogram (ECG) signals are frequently obscured with various forms of noise, including but not limited to baseline wander, electrode motion (EM) and muscle artefact (MA). The identification of noisy signals in ECG databases is challenging, particularly at scale. Many traditional signal quality assessment algorithms have been proposed, but do not generalize well. This study aimed to develop a Temporal Convolutional Neural Network (TCNN) to estimate the signal-to-noise ratio of ECG signals. Methods This study utilized a proprietary database that contained a total of 134,019 12-lead electrocardiograms (ECGs) without any machine or human added noise labels. It was assumed that this data had high signal-to-noise ratio (SNR). From each ECG, a single lead was randomly selected, and two segments were randomly chosen from EM and MA noise files. The signals and additional noise were then scaled to achieve normally distributed target pseudo-SNRs, ranging from -12 dB to +24 dB. The ECG database was split into a training set of 75% and a testing set of the remaining 25%. A TCNN with 4.4 million parameters was trained to regress the pseudo-SNR value from the raw noisy input signals. Results On the testing dataset, the mean error of the TCNN was found to be 0.33 +/- 1.73 dB. The Pearson correlation coefficient was calculated to be 0.987. Conclusion A TCNN can approximate the signal-to-noise ratio of unseen ECGs. This approach allows automated quantitive assessment of signal-to-noise characteristics of large scale ECG databases.