7Multipeak (MPEAK): Formants Plus High-Frequency Bands
Multipeak was the high-water mark of the feature-extraction era. It kept F0/F1/F2 formant tracking but bolted on three fixed high-frequency band-pass filters to recover consonant cues. It improved consonant identification — yet its formant-extraction errors in noise foretold the end of the feature-extraction philosophy.
TThe MPEAK design
MPEAK refines F0/F1/F2 by additionally extracting high-frequency information through three extra fixed band-pass filters to improve consonant perception It combines formant tracking by zero-crossing detection with high-frequency envelope detection The three added high-frequency bands are 2000-2800 Hz, 2800-4000 Hz and 4000-6000 Hz MPEAK was implemented on the Nucleus cochlear implant from Cochlear Corporation.[2008][1993]
CThe full MPEAK chain
The chain is: mic to AGC, then F0 via a 270 Hz LPF, F1/A1 via a 300-1000 Hz band, F2/A2 via an 800-4000 Hz band, plus three high-frequency envelope-detector bands at 2000-2800, 2800-4000 and 4000-6000 Hz MPEAK stimulates four electrodes; voiced segments use F0 pps and unvoiced segments use about 250 pps quasi-random For voiced sounds it stimulates the F1 and F2 electrodes plus high-frequency electrodes 4 (2800-4000 Hz) and 7 (2000-2800 Hz); the 4-6 kHz electrode is not stimulated because there is little energy above 4 kHz For unvoiced sounds it stimulates high-frequency electrodes 1 (4000-6000 Hz), 4 and 7 plus the F2 electrode; the F1 electrode is not stimulated because there is little energy below 1 kHz.[1993][2008]
TWhat MPEAK achieved
MPEAK improved consonant identification relative to the pure formant strategies It used voicing-dependent electrode selection, allocating high-frequency electrodes differently for voiced versus unvoiced segments The three dedicated high-frequency bands directly targeted the consonant information missing from F0/F1/F2 F2 was extracted with an 800-4000 Hz band-pass filter.[2008][1993]
CWhy MPEAK was superseded
MPEAK tends to make formant-extraction errors when speech is embedded in noise, degrading performance This noise vulnerability is the generic failure mode of feature-extraction algorithms that rely on estimating specific parameters The limitation motivated the spectral-maxima approach (SMSP), which extracts no explicit features and instead transmits the largest spectral peaks The shift from MPEAK to spectral-maxima marks the end of the explicit feature-extraction era.[1993][2006]
TBy the numbers
What is the most likely reason MPEAK degrades more in noise?
What did MPEAK add to the F0/F1/F2 strategy?
What weakness of MPEAK most directly motivated the spectral-maxima approach?