Ever considered using Apple's SpeechAnalyzer as the model?