Blind Speech SNR Estimation via Neural Speech Enhancement: An Enhance-and-Subtract Approach and Benchmark Against Classical Estimators

Blind signal-to-noise ratio (SNR) estimation is required when only a mixed speech recording is available. Existing estimators rely on different structural assumptions: WADA-SNR fits a waveform-amplitude distribution, NIST STNR models the low-energy mode of a short-time power histogram, and VAD-gated estimators partition the recording into speech and non-speech time regions. We propose an enhance-and-subtract estimator that repurposes pretrained DeepFilterNet3 speech enhancement as a signal-separation front end. The enhanced output is latency-aligned and least-squares scaled to the original waveform; the aligned estimate is used as a speech proxy, and the subtraction residual is used as a noise proxy. We evaluate the method on 2,560 controlled 50-second mixtures spanning eight speakers, eight noise types, and eight average-power SNR levels from -20 to +20 dB. Noise is present throughout every clip, while the speech tracks retain natural intra-utterance pauses. Using the official WADA-SNR and NIST SPQA stnr implementations, the proposed method achieves MAE 1.48 dB, RMSE 2.91 dB, bias 0.62 dB, and Pearson r = 0.980. The corresponding results are 10.03/17.99/8.22 dB and r = 0.362 for WADA-SNR, and 12.75/14.22/1.38 dB and r = -0.102 for NIST STNR. Silero VAD-SNR achieves 5.21 dB MAE and r = 0.858 on the 2,066 clips for which it returns an estimate, but fails on 494 clips (19.3%), primarily at low SNR. A separate controlled verification confirms that the official NIST executable responds to clearly separated energy regimes but compresses continuously mixed Gaussian-noise conditions into a narrow output range. We discuss this measurement-definition mismatch, the reliable operating range of the proposed method, and limitations concerning synthetic evaluation and external generalization.