Skip to content

lpform

Reads a formant's center frequency and bandwidth from an LPC pole analysis.

Use lpform to inspect resonances in a recorded voice or to use their frequency and bandwidth to control a filter. It reads the analysis loaded by lpread, rather than analyzing an audio signal itself.

Syntax

kfrequency, kbandwidth = lpform(kformant)

kfrequency, kbandwidth lpform kformant

Initialization

Create the analysis with lpanal and its -a option, which stores poles. The usual coefficient file does not work with lpform.

Select an LPC slot with lpslot, then call lpread before lpform. At initialization, lpform connects to the analysis in the selected slot. It reports an error if the slot has no analysis. Keep lpread before lpform in the instrument so the formant data follows the current analysis time on each control cycle.

Use the same orchestra sample rate as the analyzed sound. lpform uses the orchestra's sr to convert pole angles and radii into Hz.

Performance

kformant selects a pole pair, starting at 1. An analysis with 20 poles has 10 pairs. Use an integer from 1 to half the pole count. The opcode discards the fractional part and limits the index to the available range, so values below 1 select the first pair and values above the last pair select the last.

kfrequency is the center frequency in Hz. kbandwidth is the bandwidth estimate in Hz. Both outputs run at control rate. The bandwidth has a lower limit of 1 Hz and an upper limit of sr/2.

lpread orders the poles by the absolute value of their angle, so lower indices select lower-frequency pairs. These are resonances of the LPC model, not guaranteed speech formants. Their order can change as the poles move. The outputs do not include amplitude or pitch.

Changing kformant selects another pair. Changing the time pointer passed to lpread moves through the recorded analysis. The two controls serve different purposes.

Examples

The example plays the first 2.5 seconds of the manual's fox.wav speech sample and prints the first two frequency and bandwidth estimates every tenth of a second. Use it to compare the changes in a voice's sound with the resonances found by LPC.

Save fox.wav and lpform.csd in the same directory. From that directory, create a pole analysis and run the example.

csound -U lpanal -a -p20 -h256 -P0 fox.wav lpform-poles.lpc
csound lpform.csd

-a selects pole storage, -p20 selects 20 poles, and -h256 sets the analysis hop to 256 samples. -P0 disables pitch tracking, which this example does not need. Use the same Csound build for both commands. The sample and orchestra both use 44100 Hz.

Inspect LPC resonances in a spoken phrase
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
<CsoundSynthesizer>
<CsOptions>
-d -odac
</CsOptions>
<CsInstruments>
sr = 44100
ksmps = 32
nchnls = 1
0dbfs = 1

instr ReadFormants
  ; Create lpform-poles.lpc with lpanal -a before running this example.
  lpslot 0
  kTime = timeinsts()
  kResidual, kOriginal, kError, kPitch = lpread(kTime, "lpform-poles.lpc")

  ; Read two pole pairs from the analysis above.
  kFrequency1, kBandwidth1 = lpform(1)
  kFrequency2, kBandwidth2 = lpform(2)
  printks "Pair 1 %.0f Hz, bandwidth %.0f Hz | Pair 2 %.0f Hz, bandwidth %.0f Hz\n", 0.1, kFrequency1, kBandwidth1, kFrequency2, kBandwidth2

  ; Listen to the source while watching the estimates.
  aVoice = diskin2("fox.wav", 1)
  out(aVoice * 0.5)
endin
</CsInstruments>
<CsScore>
i "ReadFormants" 0 2.5
e
</CsScore>
</CsoundSynthesizer>

See also

lpread, lpslot, lpreson, lpinterp, lpanal, Linear Predictive Coding

Credits

Author Victor Lazzarini, 2009.