As a follow up to Paul Uszak's method of using a camera, I have analyzed that method for generating TRNG output, and if done correctly, it is sound, this works best with a DSLR that has a removable lens. A Sony A-6000 series camera is a good option.
There are essentially 2 sources for camera input, visible light and radiation that causes the sensor to think it is getting visible light. The easiest way to get radiation input is to take the lens off, put on the lens cap, and point the camera at a radiation source, preferably a beta source, such as uranium ore or an isotope disk. A pure alpha emitter will possibly be blocked or not generate a signal, and a gamma emitter might damage the sensor and will probably not be detected anyway.
If going for radiation, my suggestion would be Thallium-204, which is a beta emitter, combined with an alpha/beta emitter such as Uranium ore or metal. If nothing is detected with a long shutter speed, try taking off the lens cap.
Once you have adjusted the ISO/shutter/f-stop/focus settings to get a good particle activity pattern, start snapping pics. A 24 megapixel camera picking up 1% of the pixels with particle activity will get you at least 24KB worth of entropy after processing per image, which can be converted very safely into usable bits using Keccak as an entropy conditioner.
If you are using visible light with ISO noise as your source, I would go with a clear night sky with the focal point as close as possible to blur the stars out, this gives you photons at a pretty constant rate but sends them all over the sensor. Take each image with the lens pointing in a slightly different location. Use the smallest aperture/highest f-stop.
You can also use a radiation source for light, apply phosphorescent paint to the inside of cardboard box that will be sealed to block all light, and adjust the focus point manually on the paint. Then place the emitter in the box with the camera, and it will excite the paint so it can be detected by the camera. I think you may be able to mix Uranium Dioxide powder directly into glow paint, but high output beta emitters at a distance will give a more random pattern.
Conditioning is required to eliminate bias and convert the input, which may be extremely sporadic in terms of bit rate, into a usable random bit source.
My recommendation for converting the TRNG data source into usable bits would be a pair of Keccak instances both operating in an unbalanced duplex mode. If you have an image from a 24MP DSLR, capturing 6000x4000 pixel images, break that into 96 500x500 pixel sub-images if you are getting a full frame image of activity. Absorb 8 sub-images to set the initial state. For each additional 4 sub-images absorbed, squeeze 288-bits. You will get 6336 bits per image. Process 2 images in parallel using this method and XOR the outputs together, resulting in 6336 bits from 2 images. At 47 bits per 10 letters, you get 1348 letters per 2 images, so 6 images gets you the required 4000 letters.
Obviously, if the camera is lower resolution you will need more images. If you are not getting activity at the edges of the image, ignore all edge blocks and use only the inner 60, which gives you 3744 bits or 796 letters. You will need to do the math yourself based on the resolution of the camera, but 1000x1000 or 500x500 is a good sub-image size to use for this application, 500x500 makes it easier to omit the edge blocks without losing most of the image. Smaller blocks also mean more mixing of the image into the Keccak state.
Breaking the entropy conditioning down, output bits is 13.2% of the input, or a ratio of 7.57:1. This assumes a minimum entropy of 1 bit per 1000 pixels, which is extremely conservative, you will probably be getting at least 100. Since you can afford to be conservative with this method due to the low number of letters required, that is just fine. If you want a consistent bitrate into the Keccak instances for auditing and performance purposes, preprocess the sub-images with a 512-bit hash function first.