A phonetic vocoder for finnish

A phonetic vocoding system for Finnish is described. This very low bit rate speech coding method is a subclass of segmental vocoding. The proposed system utilizes speech recognition based on a HMM system to phonetically segment and label the input speech. The fundamental frequency is estimated using a robust pitch tracking algorithm. Speech reconstruction is carried out with a conventional LPC speech synthesis algorithm. This approach makes possible a higher-level description of the speech than usually obtained in coders, thus enabling e.g. speaker transformation relatively easily. The coder has a high computational complexity and requires a delay of several hundreds of milliseconds, but the obtainable bit rate is under 500 bits/s. The speech intelligibility of the current implementation is high but the overall speech quality is relatively poor.