Summary

• Robust EM algorithm, giving good frequency estimates
• Weak partial's frequencies are constrained by the stronger ones of the same source, using harmonic structure
• Credibility measurement via relative interference of the spectral envelope
• Gives highly accurate estimates when not too complex
• Slightly degraded when highly overlapping but still better than initial estimates via peak-picking
• Does not cope very well with time-varying parameter signal but still gives decent estimates
• Stable in all cases
• Relies heavily on initialization
• Still a problem in amplitude estimation. Solving in a frame is impossible. Need temporal information or other priors
• Future work includes other form of constraining, credibility measurement (probabilistic, perhaps), solving amplitude inaccuracy problem using other info and extension to non-stationarity model.