I work on voice recognition too. Yes, very brilliant query about sound perception that I can clarify and not answer completely. Most of the text belongs to DSP SE forum, i.e. Voice rec, MFCC, filter, BP, audio processing. You may have confused the readers except for the image and the final paragraph.
I had to rephrase it to clarify it to myself:
Why are humans unable to hear pitches on a straight maths scale, they hear logarithmically? We hear according to a mystery ratio called MEL distribution, a well studied summary of many subjects perception of pitch, averaged to:
0.01% Humans have Perfect pitch, 95% are relative pitch, and 5% are tone deaf. MEL varies by the type of listener. Tone deafness is hereditary, ( info about biology and brain structurehere) and the rest of humans probably follow a gradual distribution of poor to good pitch differentiation, and they also hear amplitudes differently.
As you can see from the wiki page, there is no explanation for MEL theory, it's either neurological, a soundwave energy balance, or because of the ear anatomy.
To answer the title question, your graph shows humans have a continuous response to frequency spacings through all their perceived sound. However volume tests show amplitude sensitivity peaks at about 3kHz, which you can see on inaccurate amplitude perception and equal loudness contour graph.
MEL studies of Steinberg are stagnant since Volksmann, same as Donald D Greenwood suggested in 2009, they overgeneralize and the field needs to be developed a lot further under a different study name than MEL scale, which isn't english.
Why can they hear voice well under 1000Hz?
At least 1/2 of the sound energy of voice occurs below 1000Hz, same reason bass requires a lot of energy.
So humans can hear sub 1000Hz sounds from further away, but the more sound data you have the clearer the voice is, and computer should use nearly all of possible human voice frequencies to measure formants and N/M consonants to best accuracy. Sub 1000Hz may contain enough vowel subharmonic frequency bands at different spacings to represnt the vowels of sound basically, but it is less clear and realistic if the voice is filtered at 1000Hz, the subharmonics of AEIOU go multiple octaves above middle C male/female voicebox peak, near 440Hz.
The sub 1000Hz components are generated by the voice box, which changes tone at a slow rate, whereas the 2000-3000Hz sounds, i.e. Frictives, S,F, and N aspirated P, k, don't require very as complex sound processing to interpret (for digital sound filter processing)
The 625-1000Hz pivot component of the equation doesnt appear less straight under 1000Hz if the graph is from a parametric frequency view.
To resolve your query, Biologists would want to know:
- 1/Structure of the ear as a reason for sound perception imbalance
- 2/Child learning and adult learning of complex pitch tasks
- 3/Brain structures used in pitch recognition
- 4/Evolution of ears and psychoacoustics, reason for a hearing peak at 3k which is the same as rustling leaves
- 5/Physics energy and mathematical reason for MEL distribution (ask on DSP SE)
As a Voice rec researcher, you should think outside of the box. Computers don't hear like humans. Frictives like F and S have peaks around 3000-4000HZ, they are the highest and easiest to recognize and similar in all accents. Formants from vowels can peak from 400 to 500HZ roughly in men and women, at 700hz for children.
Vowels are vague and vary by accent are are the least precise element in voice-rec. That's why voice recognition hinges a lot on formants and accents which have energy peaks below 1000HZ, and timbral graphs which change wildly wether it's AEIOU sounds.
As a side not for voice-rec,
English is tougher to recognize than French because of the stress time, swingy nature and the 2-3-4 vowels per syllable, and perhaps that's why French have lost the wild traditional accents faster (I know I speak both since I was 0). Voice rec of French and other syllable time languages is easier.
I don't know why MEL is used at all in voice-rec filters. Perhaps because people use the same intonations as they hear from MEL, so the PC must know too, else because the programmers don't know enough about real sound physics, and have implemented it as a rough power-balancing graph for filter results? else some other physics reason. Other power-related filter shapes should work fine, as DSP SE if this has a known physical basis of if it uses human psychoacoustics for digital audio inaccurately:
mel-frequency cepstrum (MFC) is a representation of the short-term power spectrum of a sound, based on a linear cosine transform of a log power spectrum on a nonlinear mel.
Humans don't need perfect pitch for anything except music. Relative pitch, we nearly all have ok, and it makes sense because biologically relevant sounds, animal calls, physical signals of water, trees, voices, add together to make a continuous freqency spectrum that we have to measure all the frequencies by a similar rating of importance to interpret best. Voice contains lots of micro-expression subtones that represent clear psychological states that humans benefit to hear, and rustling leaves and other sounds all should biologically benefit from a high relative pitch clarity.