fmanal - formant estimates from speech waveforms


fmanal (-w window) (-s step) (-v|-t offset) (-f fixtxsize) (-p numpoles) (-i item) file


fmanal performs a windowed LPC analysis on sections of a speech waveform and solves the predictor polynomial to obtain spectral peak (formant) estimates. Analysis windows may be of fixed length with a selected overlap, or may be positioned by means of a TX item (in which case the unvoiced periods are analysed according to the fixed length parameters). Pre-emphasis, a Hamming window and the autocorrelation method is used for all analyses. If a Tx item used to position analysis windows, the window size defaults to the pitch period. However, an explicitly given fixed window size (-f option) will over ride the default.

Options and their meanings are:

-I Identify program and version number.

-i item Select input item number.

-w window Select analysis window size in milliseconds. Default: 20ms.

-s step Select analysis window step size in milliseconds. Default: 10ms.

-f fixtxsize Select fixed analysis window size in milliseconds for excitation synchronous analysis.

-v Treat all frames as voiced. Default is to use autocorrelation pitch detector and adapt analysis to voicing decision. Voiced frames are pre-emphasised, unvoiced frames use half as many poles.

-t offset Select excitation synchronous analysis with supplied offset in micro-seconds between tx markers and start point of analysis window. Negative is earlier, positive later. Windows less than 20ms are treated as voiced.

-p numpoles Specify number of poles to use in LPC analysis. Default is (2+sample_rate/1000).


1.xx Any speech waveform (10kHz recommended)

3.xx (Optional) Excitation markers.


12 Formant estimates.


1.0s - Mark Huckvale
1.2s - Stuart Rosen (modifications)
1.5 - Mark Huckvale

Fri Jul 09 14:54:35 2004