Conclusion

After an interesting sonic journey I have found that I can create a great robot sound using the phase vocoder. However it is not the sound I was searching for. I was anxious to get the classic robot sound created by old vocoders which modulated the voicing information form one signal with the frequency information of another. This voicing and pitch separation is very different from the magnitude and phase separation achieved with the phase vocoder. It was not clear to me until the tail end of the project just how important this distinction is.

In my original implementation of the phase vocoder, the filter bank, I ran into issues of filter design. One paper I read suggested using 6th order Bessel Filters. These filters have a linear frequency response and great transition bands for the task at hand and are only 6th order. However when I translated these into the digital real using remez they bacame very large and they no longer had acceptable phase characteristics. This led to investigations into using digital FIR filters. These filters became impractical in a production environment since the ones I calculated were on the order of 700. This number of course could be greatly reduced of course by experimenting with the transition bands and ripple. However it was dangerous to at that point since I would have more trouble eliminating variables while debugging and testing my vocoder early on. Luckily I was led to the Fourier Transform interpretation before I wasted more time on the filter bank.

Another important issue I had with the filter bank was being side tracked by trying to get the classic vocoder sound. I spent hours playing with the broken up signals trying to get an interesting effect. At first I thought my vocoder wasn’t working. In the end, that wasn’t the case, I was just trying to use it to do something it wasn’t designed to do. My naive understanding of the channel vocoder led to misconceptions of the phase vocoder. This experimentation time could have been better spent investigating and implementing more elaborate pitch detection and shifting techniques. In my harmonizer and chorus effects for example, it would be best to use up to three peaks to use as input for the harmonizing and chorusing. In terms of the pitch shifting it would be interesting to use a technique that didn't require resampling.

knowing that the phase vocoder cannot recreate the modulated robot voice of the channel vocoder, more energy should have been focused i other areas. For instance, improving the effects using the phase vocoder or utilizing formant or other linguistic technics to simulate the voiced speech spectrum. This could have led to some interesting variations on the classic channel vocoder sound.

Overall the phase vocoder was successful in generating multiple effects, including a very solid time scale operation and robotic voice. Being able to explore many different effects using similiar techniques what a great exercise. Having explored two very different implementations of the same idea gave me some good insite into DSP techniques as well as some positive and negative aspects of both analog and digtal processing. Now that these effects are built they could be used to create some interesting audio clips or music. There are lots of creative uses for these musical operations.

future work

It would be very easy to expand on the items in this project. Future work based on this phase vocoder would include the following. Implementing the pitch detection techniques described in the paper by Jean Laroche and Mark Dolson and generating more realistic harmonizing and chorusing effects. Capturing the formants of the signal and experimenting with modulating the formants to get interesting audio effects. Future work could also lead to readdressing this robot voice modulation but this time using a formant vocoder which attempts to split this sound in formant information.