lpanal

Linear predictive coding (LPC) analysis utility

The lpanal utility performs linear predictive coding (LPC) analysis on sound files. LPC models the vocal tract as an all-pole filter and is particularly effective for analyzing and resynthesizing speech and vocal sounds.

Syntax

lpanal [flags] infilename outfilename

Arguments

• infilename - Input sound file to analyze
• outfilename - Output LPC analysis file

Options

File and time selection

Option	Description	Default
-s <srate>	Input sample rate	From header or 44100
-c <chnlreq>	Channel to analyze	1
-b <begin>	Start time in seconds	0.0
-d <duration>	Duration to analyze in seconds	To EOF

Analysis parameters

Option	Description	Default
-p <npoles>	Number of poles for analysis	34
-h <hopsize>	Hop size between frames in samples	200

Pitch tracking

Option	Description	Default
-P <mincps>	Lower limit for pitch search in Hz (0 disables pitch tracking)	70
-Q <maxcps>	Upper limit for pitch search in Hz	200

Additional options

Option	Description
-C <string>	Comment field for LP header
-v <verblevel>	Verbosity level (0=none, 1=verbose, 2=debug)
-g	Enable graphical display of results
-a	Use alternate (pole) file storage format
-n	Use Durbin method for linear prediction
-- <fname>	Log output to file

Examples

Basic LPC analysis

lpanal input.wav output.lpc

Analysis with custom pole count

lpanal -p 50 voice.wav voice.lpc

Speech analysis with pitch tracking

lpanal -p 40 -P 80 -Q 300 speech.wav speech.lpc

High time resolution analysis

lpanal -h 100 -p 30 input.wav output.lpc

Analysis without pitch tracking

lpanal -P 0 input.wav output.lpc

Analyze specific time segment

lpanal -b 2.0 -d 5.0 input.wav output.lpc

Verbose analysis with comment

lpanal -v 1 -C "Male voice analysis" voice.wav voice.lpc

Output file format

The LPC file contains for each analysis frame:

• Filter coefficients: Predictor coefficients modeling the spectral envelope
• Gain: Overall energy/gain for the frame
• Pitch: Fundamental frequency (if pitch tracking enabled)
• RMS: Root mean square amplitude

Choosing analysis parameters

Number of poles

The pole count determines the filter order:

• 20-30 poles: Minimal formant resolution, fast analysis
• 34 poles (default): Good balance for speech, captures 16-17 formants
• 40-50 poles: Higher spectral detail, better for singing voice
• 50+ poles: Maximum detail but may overfit

Rule of thumb: Use approximately samplerate / 1000 poles for speech.

Hop size

Controls time resolution:

• 50-100 samples: Very fine time detail, more data
• 200 samples (default): Good balance (4.5ms at 44.1kHz)
• 400-500 samples: Coarser time resolution, smaller files

Frame size is automatically set to 2 × hopsize.

Pitch tracking range

• Male speech: 80-250 Hz
• Female speech: 150-400 Hz
• Singing voice: 60-1000 Hz (or wider)
• Musical instruments: Disable with -P 0

Technical details

Linear prediction methods

Autocorrelation method (default):

• Stable, always produces minimum-phase filters
• Guaranteed to find a solution
• Standard method used in speech coding

Durbin method (-n flag):

• Alternative recursive algorithm
• Computationally efficient
• May give different results for some signals

Frame analysis

For each frame, lpanal:

1. Extracts a windowed segment of audio
2. Computes autocorrelation function
3. Solves linear prediction equations
4. Extracts filter coefficients and gain
5. Optionally performs pitch detection

Using LPC files in Csound

LPC analysis files can be used with Csound opcodes:

• lpread - Read LPC data from analysis file
• lpreson - Resynthesise using LPC filter
• lpfreson - LPC resonator with frequency control
• lpslot - Store LPC data in slots
• lpinterp - Interpolate between LPC frames

Related utilities

• atsa - ATS sinusoidal analysis
• pvanal - Phase vocoder analysis
• hetro - Heterodyne analysis