Formant tones: weak components at formant frequencies actively generated in the vocal tract

The authors have discovered the presence of weak sinusoidal components located at formant frequencies in voiced speech segments. They refer to them as formant tones and propose a method to estimate and track the frequencies of these formant tones. These sinusoidal components are often located close to the strongest harmonic component under the formant peak and closely track the respective formant peaks with time. The formant tones seem to provide crucial incremental information required precisely to identify and track the formant location. These tones are not easily discernible using standard short-time Fourier transform or linear prediction methods. However, when the speech signal is decomposed into its harmonic components, tell-tale evidence of these formant tones is visible in their envelopes. Then the strong harmonic components are removed by high-pass filtering the envelopes. The high-pass filtered envelopes are subsequently used to identify the frequency tracks of these tones. Since the formant tones are non-harmonic, they must originate in the vocal tract as a reaction to the impinging glottal pulses. These tones appear to be angle-modulated by the glottal pulses.<>