Effects of noise removal from speech recordings

Speech recordings often have ambient background noise from water, fans, animals, and humans. This noise can make transcription difficult and has an unknown effect on phonetic measurements. Denoising tools may help, but researchers have not studied their impact on phonetic measurements. The effects of aggregated ambient noise overlays (town, dogs, crickets, water, fan, and geese) at five signal-to-noise ratios (SNRs) (20, 15, 10, 5, 0 dB), using various denoising systems (none, seewave, Python's noisereduce, Audacity, iZotope, and FFmpeg's RNNoise) were compared to gold-standard recordings from Sanker, Babinski, Burns, Evans, Johns, Kim, Smith, Weber, and Bowern [(2021). Language 97(4), e360-e382]. The resulting audio was transcribed using the Montreal Forced Aligner (MFA) and analyzed with PRAAT. Using two one-sided equivalence tests bounded by published just noticeable difference values, results show that MFA and PRAAT are robust against unaltered ambient noise for consonant duration, fundamental frequency, and spectral tilt, and vowel duration and jitter to 0 dB SNR, and center of gravity (COG) to 5 dB SNR. F1 results benefit most from Audacity, F2 from noisereduce, and COG from RNNoise, with some other tools helping in specific circumstances. Harmonic-to-noise ratio measures require low background noise and cannot be repaired through denoising. Denoising tools strongly influence phonetic measures, and their use should be declared alongside other recording metadata.