This paper describes a new method for estimating formant frequencies. It operates in two phases. The first phase, which is similar to a technique developed by Talkin [JASA, vol. 82, S1], finds optimal formant track estimates by imposing frequency continuity constraints using Dynamci Programming (DP). DP is used to select a mapping of candidate frequencies to formant frequencies in oral sonorant regions based on the minimum cost from all possible mappings. The second phase performs a series of postprocessing steps to make formant estiamtes more robust and accurate and extends the formant estiamtes into nasal and obstruent regions. Performance statistics comparing the formants obtained with this technique with a set of reference formants using 34 sentences randomly selected from the TIMIT database shwos our algorithm gives excellent results when he formants are among the candidate frequencies.
展开▼