Answers are grounded in this paper and cite their sources.
A loudness leveler for live voice that must not delay the signal cannot look ahead, and a neural voice detector that could tell it what is voice works on frames and reports late. We describe Themis, a leveler whose audio path is one per-sample gain without latency and whose voice detector is a frame-based recurrent network (RNNoise) running beside that path; two instantaneous rules cover the lateness of the network's verdict. We evaluate it with a gain-recovery protocol that gives the loudness of the voice and the level of the background at the output of any gain device, on 87 public sessions of English, Mandarin and Korean speech and of singing, at four input levels, five noise conditions, level steps and long pauses, against WebRTC AGC2, the FFmpeg filters loudnorm and dynaudnorm, a fixed compressor and its own level-only variant. On speech, Themis holds the voice within 0.63 LU of the target for inputs up to 18 dB below it, on a par with the look-ahead normaliser loudnorm (0.68 LU), and reduces the utterance-level spread from 3.3 to 1.6 dB. The network's contribution is confined to long pauses, where the background is raised by 2.7 dB instead of 7.8 dB at the end of pauses of 10 s or more at 20 dB SNR, and to SNR of 10 dB and below. On singing the speech-trained network fails: its voice-share estimate takes sung notes for noise and the voice overshoots the target (mean absolute error 2.8 LU against 1.0 LU for the level-only variant). Objective quality predictors do not favour the processor in noise. No listening test was conducted; the claims are limited to measured loudness, gain behaviour, latency and cost.
No takes yet. Share an insight, caveat, or question.
YUSUHAO (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: