Jaffe Synthesis Evaluation part II criteria 6 & 7

Jaffe Synthesis Evaluation part II criteria 6 & 7

1. how EFFICIENT is the ALGORITHM

   • EXTREMELY IMP. - FOCUS OF MY DISSERTATION
     • DETERMINES:
       • 1. REALTIME
         • NOS. VOICES AVAILABLE ON GIVEN HARDWARE
       • 2. IN NON-REAL-TIME CONTEXTS -
         • TIME OF WAITING FOR RESULTS OF COMPUTATIONS
           • IN A FIXED AMOUNT OF TIME TO COMPLETE A PIECE AN EXTREMLY LONG FEWER ITERATIONS == LESS-REFINED RESULTSTURN AROUND TIME =
   • QUOTE: "determining the efficiency of an algorithm is more complicated than it might at first appear. It is not merely a matter of comparing processing benchmarks. Numerous aspects of a technique and its implementation come into play,"
     • The aspects of the GS techniques and their implementation that come into play can be divided into 3 categories:- a memory, b processing requirements c control stream heaviness
       • IaMEMORY REQUIREMENTS
         • SINE E.G.Amem - WAVETABLE SYNTH
           • Mathews '69
         • VARIABLE MEM REQ. Changes with the parameter values
           • Vm1 Karplus-Strong (plucked str) (waveguidebased MODELLING): MORE MEM LO Pitch LESS MEM HI Pitch
       • Ib PROCESSING DETAILS
         • SINE.G.Bproc - ANALYTICAL COMPUTATION
           • Alt. techniques for sinewave generation
             • EGB1proc
               • marginally stable 2-POLE FILTERS
             • EGB2proc
               • EVAL. of COMPLEX PHASOR
                 • Gordon, Smith '85
             • EGB3proc
               • WAVEGUIDES
                 • Smith, Cook 92
         • II (b) PROCESSING POWER
           • 1. quite COMPLEX"
           • 2. depends on the PROCESSOR ARCHITECTURE
           • poor EXPENSIVE:
             • EGIIB
               • Finite Element Modelling (numerical INTEGRATION of DIFFERENCE EQUATIONS that Descrb. Masses & Springs
                 • BUT wellsuited to: ARRAY - PARALLEL PROCESSORS
           • good EFFICIENT
             • 1. active code fits within the CACHE of a RISC CHIP RUNS FASTER than an active code set that OVERFLOWS THE CACHE
               • AFreed, Rodet and Depalle 93
             • 2. With MINIMUM CHANGES in PROGRAM FLOW
               • well supported by:
                 • a. DSP processors
                 • & other b. HEAVILY PIPELINED ARCHITECTURE
             • PURPOSE BUILT HARDWARE INCREASES ENORMOULSY THE EFFICIENCY OF A TECHNIQUE
             • GOOD
           • B Some techniques have a PROCESSING REQUIRMENT that CHANGES with the PARAMETER VALUES
             • Time domain implementation of CHANT
               • more EXPENSIVE as FREQ RISES
                 • MORE PITCH PERIOD/SEC.
                   • MORE ADDONS and more TABLE LOOKUPS PER OUTPUT SAMPLE
             • Vm2 N harmonic partials in Trainlet synthesis - brightest timbre (chroma) = bandlimited pulse, softer timbre (low pass filtered, ie higher partials attenuated
               • with MORE MEM/MORE PROC = NARROWER PEAK with 32 Harmonics than with 8 - changes behaviour of the technique
           • Formant Wave Function FOF
             • synth. by adding overlapping vocal tract IResp.
       • c CONTROL STREAM ATTRIBUTES
         • density/heaviness of control stream
           • Problem esp. for RT impl.
             • may need 2 processors - then expensive processing to get data from control processor to the sound procesor
             • even wth just 1 processor - is their enough DISK SPACE/BW for the cntrl stream data ie is control stream density > RAM
         • Dispersion pattern:
           • sys. FAIL wth sporadic, CLUMPED, LARGE DENSE BURSTS of parameter update mess,
             • e.g. ENTIRE AMP ENV chngs with note data, at be of each note
           • more easily manageable: REL STEADY STREAM OF WELL-SPACED FAIRLY DENSE CNTRL MESS.
             • More EFF. feed ENV DATA BY BREAK POINT THRU THE NOTE 1 POINT AT A TIME
       • TRADE OFF
         • well known axiom
         • A single period of a waveform stored in memory
           • EGA1mem
             • use a non-interpolating 'drop-sample' oscillator
             • stored in memory in a HUGE TABLE
             • Greater Memory less processing
           • EGA2mem
             • stored in memory in a SMALLER TABLE
             • a more expensive OSC that interpolates between samples in the table
             • Optimised Mem - but MEM < EG1 PROC> EG1
           • EGA3mem
             • Precomputed resultant. WAV: ƒ WAVE * A ENV
               • MAX MEM MIN RT PROC

2. how SPARSE / DENSE is the CONTROL STREAM

3. what CLASS of SOUND can be represented

4. what is SMALLEST possibel LATENCY

5. do ANALYSIS TOOLS exist

6. 1 to 5 relate to parameter behaviour, physical, intuitive, wellbehaved, perceptible changes, sound ID robust e.g. sounds good on more than 1 pitch/resonance of the instr.