Speech formant trajectory estimation using dynamic programming with modulated transition costs

A new algorithm to track automatically speech formant frequencies have been developed. Dynamic programming is used to optimize formant trajectory estimates by imposing appropriate frequency continuity constraints. The continuity constraints are modulated by a stationarity function. The formant frequencies are selected from candidates proposed by solving for the roots of the linear predictor polynomial computed periodically from the speech waveform. The local costs of all possible mappings of the complex roots to formant frequencies are computed at each frame based on the frequencies and bandwidths of the component formants for each mapping. The cost of connecting each of these mappings with each of the mappings in the previous frame is then minimized using a modified Viterbi algorithm. Two sentences spoken by 88 males and 43 females were analyzed. The first three formants were tracked correctly in all sonorant regions in over 80% of the sentences. These performance results are based on spectrographic analysis and informal listening to formant-synthesized speech.