Voice Banking with Model Talker

Although that I'm not showing any definitive changes to my voice at the moment (that I cannot attribute to paranoia anyway) and knowing the progression of ALS, I decided that it would be wise to create a voice bank so that when that unfortunate day does come.  I'll still be able to communicate using speech to text technology and it will sound somewhat like me.

Creating a voice bank is best done when your voice is still clear and strong. Definitely something that should be done before symptoms start affecting your speech patterns. Reason being is any speech impediment you may develop from ALS will consequently be recorded in the voice bank. Not to mention that the recording process is actually quite long. It can take quite a significant amount of time and effort to get done properly. And even not showing symptoms, it was at times quite fatiguing.

I used a software called model talker which required me to record about 1600 phrases. Most of the phrases were quite simple, with many references to the Wizard of Oz. There were however some reasonably challenging tongue twisters. This software then takes all the phonemes and uses them to create a text to speech file. This file can then be used within the Windows text-to-speech systems or (so I'm told) used with other hardware such as Dyna Vox.

say hello to my little voice bank

Model talker is presently in beta phase. So you can expect the possibility of some instabilities at times. Personally I did not counter many of these. The issues I did have had to do with setting up the recording in my own environment. For some reason any microphone I tried on my home computer was also recording a white noise or hum. After trying multiple microphones I could not isolate the cause. Best I can figure is this EMF field somewhere in my house, or noise from the power bar which is disrupting my audio recording (for example, check this video. The noise in my recording was not quite so pronounced enough to disrupt the quality of the recording). This is not something I could blame on model talker. I ended up having to record my voice and my rehabilitation facility. It was a bit inconvenient to go there so frequently, but I felt it was a necessity to create the voice bank. Recording their went quite smoothly.

The people at model talker were extremely helpful and responsive. Whenever there was a problem on the initial calibrations they would get back to me very quickly. Also once I submitted the final files, I was expecting it to take up two weeks to get back. They had my full voice bank ready for download by the next afternoon.

On a side note, if you need to make a recording of your synthesized voice as I did for this blog and you're using speech recognition for dictation… Be sure to turn off the microphone for speech recognition. I forgot to do so the first time and the speech recognition was actually dictating from my synthesized voice… LOL!

This download included two versions of playback software which could be used in addition to the Windows text-to-speech functionalities. So far, my best results seem to have been from using model talker 2. I was a little dissatisfied with the initial playbacks and found my voice didn't sound enough like me to make the effort worthwhile. In particular there seemed to be too much of a gurgle. However after tweaking some of the settings (unfortunately model talker 2 didn't include any documentation for this) I found that setting the phonetic target cost down to about 20 made for a significant improvement in the gurgling issue. I'll probably be able to improve on it further once I figure out more about the other settings.

Here, listen for yourself ;-)

• Pike synthesized voice sample
• Pike synthesized voice chaos theory

can you hear me now?

However, being beta I did have some difficulties in recording the samples. The record option fails and changes to the synthesis settings do not seem to be retained between executions. That being said, it is presently free for you to create your own voice bank.