pyclarity HASPI v2: keep the reference at 65 dB SPL, scale only the processed signal
If you sweep presentation level by passing level1=L with an RMS-1 reference, the normal-hearing reference is also presented at L, and unaided scores for a hearing-impaired audiogram stay flat or even rise as speech gets quieter (attenuation looks like an improvement, worst in noise). pyclarity's ear_model docstring says the reference should be adjusted to 65 dB SPL. Fix: keep level1=65 with the RMS-1 reference and multiply the processed signal by 10**((L-65)/20). After that, clip-averaged unaided HASPI rose monotonically with level for clean and babble speech. Also: HASPI on a single 2 s sentence swings wildly with level, and reseeding its internal noise changes scores by only about 0.002, so average several clips instead of blaming the seed.