Singing voice synthesis
A singing voice synthesiser with controlled expressivity was created using Max MSP. The first aim of the project was to synthesise a tenor singing voice who sings the very last bars of Puccini "Nessun Dorma".
The second goal of the project was to be able to smoothly change the characteristics of the voice from a robotic voice, to an espressivo voice. The basic design of it is a formant synthesiser. Noise and saw-tooth signal generators were used as a source, and they were filtered using bandpass filters.
An analysis of a real voice singing the lasts bars of "Nessun Dorma'' was carried out using Praat. The results of this analysis were the key for a perfect synchronisation of the parts, as well as for the naturalness parameters of the espressivo voice. To synchronise it, the events within the Max patch, phonemes lengths, and their position were exported from this analysis.
The primary goal of the project was to be able to change the expressivity of the voice using a single slider. Therefore, an interpolation between the values generated by both espressivo and robotic parts was needed. The interpolation was made before changing any parameter. Thus, the first point to interpolate could be for example the raw frequency value from the espressivo qlist, the Point 2 could be the result of the vibrato generator in the robotic part, and the resulting interpolated value would be the actual value. Next, this value changes the saw generator object. The same behaviour is happening with the bandpass filter central frequency.
Finally, the user interface for this project has been designed to be easy to use. The main component is the big horizontal slider which controls the espressivo and robotic parameter of the synthesiser. At the left part of the screen, a big button starts and stops the system, whereas at the right part of the screen, three vertical sliders let the user mix the different audio components.
The second goal of the project was to be able to smoothly change the characteristics of the voice from a robotic voice, to an espressivo voice. The basic design of it is a formant synthesiser. Noise and saw-tooth signal generators were used as a source, and they were filtered using bandpass filters.
An analysis of a real voice singing the lasts bars of "Nessun Dorma'' was carried out using Praat. The results of this analysis were the key for a perfect synchronisation of the parts, as well as for the naturalness parameters of the espressivo voice. To synchronise it, the events within the Max patch, phonemes lengths, and their position were exported from this analysis.
The primary goal of the project was to be able to change the expressivity of the voice using a single slider. Therefore, an interpolation between the values generated by both espressivo and robotic parts was needed. The interpolation was made before changing any parameter. Thus, the first point to interpolate could be for example the raw frequency value from the espressivo qlist, the Point 2 could be the result of the vibrato generator in the robotic part, and the resulting interpolated value would be the actual value. Next, this value changes the saw generator object. The same behaviour is happening with the bandpass filter central frequency.
Finally, the user interface for this project has been designed to be easy to use. The main component is the big horizontal slider which controls the espressivo and robotic parameter of the synthesiser. At the left part of the screen, a big button starts and stops the system, whereas at the right part of the screen, three vertical sliders let the user mix the different audio components.
Download the patch
| |||