Hard to give an easy answer, as this touches multiple areas, so here's just an overview of what CSM and speech is about.
The Yamaha FM Operator Type L (OPL)(2) (*1) is as great because it does not use some analogue circuits to create some sounds, like for example the C64's SID, but calculates sounds digital only to be turned into an analogue signal at the very end. This abstract approach - nowadays, with fast processors standard, allows the creation of a high number of operators, each acting like an independent sound channel. For CSM it's important that these can be operated as 9 generators of two sine signals (*2) each.
In CSM mode these are used to 'recreate' the complex sounds of speech as one theory, the sine wave synthesis, describes them: as a collection of pure sine waves.
The idea is, that a section of speech can be described as a set of sine waves, which would reassemble the frequencies the various parts of our speech tract creates for that sound. Since each sine can be described as a single number, one would only need a very low amount of data to encode such a section. At the same time, since the resulting signal is a rather complex one, allowing these sections to be way longer than what is needed for a full figured digital recording.
For somewhat acceptable (phone) quality a straight digitalization might need some 6k samples per second or about the 6 kByte/s. With sine wave synthesis each section could be as long as 20-50 ms while still creating acceptable speech quality. This means only 20-50 sets of sine arrays are needed per second. So, if for example a sample can be knocked down to 20 sine waves (or less), a second of speech is no longer 6k but only 0.4 to 1 kByte. Sounds like a great compression rate, doesn't it?
When it comes to encoding the method is called Linear Predictive Coding (LPC), which is all about encoding and recreating speech like sound using a low number of formants - in case of the OPL, these formants are simply the sine waves we talked about before. LPC is base of various codecs, used for digital phone services like various mobile or VoIP. Of course these use way more sophisticated methods than just selecting the 10 highest sine waves of a section.
And yes, LPC is what TI used as well for their speech chips, from Speak&Spell all the way to the 99/4 Speech Synthesizer. Although, their method of creation was even compactified than the use of sine waves.
For usage of the OPL, CSM means that believable speech can be created by feeding one set of sine values to all 9 generators and have it repeat them for maybe 50 ms before feeding the next table. A task quite solvable already for 8 bit micros.
*1 - OPL and OPL2 are for most parts the same, the OPL2 just added some more wave forms.
*2 - The whole base for the OPL is a ROM with data for a quarter sine wave