Pitch Perfect

I/O

We used javax.sound standard libraries to read and write audio WAV files, capture input from a connected microphone, listen for MIDI input from a keyboard, and playback output in realtime to the speakers.

The FFT

We coded our own implementation of the FFT, a divide-and-conquer algorithm to compute the DFT of a signal efficiently. This allows us to analyze and manipulate the signal's fundamental frequency and harmonics directly, taking into account phase information as well.

Pathway

We process our input (from a file/microphone) in overlapping chunks of a fixed size. After first FFT-ing the chunk, we detect the source pitch present. We then determine the target pitch for the chunk by either detecting it from a separate input file or acquiring it in realtime using MIDI. Using a technique known as phase vocoding, we can then change from the source pitch to the target pitch. Reassembling the chunks yields the pitch-altered signal.

Pitch Detection

First we determine if the sound is voiced or unvoiced. In speech, voiced sounds come from the vocal chords, and unvoiced comes from the lips, tongue, and other parts of the mouth. If the sound segment is unvoiced, we say that the pitch of current segement is equal to pitch of previous segment.

Next we use algorithms like AutoCorrelation and Cepstrum Biased HPS to determine various possible pitches in the sound segment.

Finally we weigh the possible pitches, taking into acount standard deviations from the mean and possible artifical doubling or halving in the pitch detection algorithms. We determine the pitch based on maximum weight

Phase Vocoding

In order to shift pitches with minimal distortion, we implemented a phase vocoder. This involved using the FFT to get frequency bins, finding peaks in those bins, and shifting those peaks to alter the pitch.