This sounds like a security breach that would defeat row-level encryption.
Soundex is an algorithm for indexing words by their approximate sound. You can think of it as a lossy compression approach. As a lossy transformation, it is not reversible in the sense that you can directly extract the original input. However, it is very much reversible in the sense that you can generate plausible inputs. Since the Soundex index would disclose a lot of information about the plaintext
- an attacker could use this to crack the encryption far more efficiently
- an attacker might not even need to crack the encryption to make inferences about the plaintext content. Indeed, that you are creating this index means that you want to be able to search the plaintext contents without having to perform the decryption first.
So storing encrypted text next to a Soundex-index of that text will likely break your security goals.
In contrast, such a Soundex index would not be a problem if both the actual text and the index are given the same protection level, e.g. if your encryption works at the disk or database level.
It might be a good idea to take a step back and think about your actual security goals and your threat model. What is the encryption supposed to achieve? Which attackers with which capabilities are you defending against? Is row-level encryption even necessary or helpful in that threat model? Do you really need an unencrypted Soundex index? Some problems cannot be solved though – there is often a security–convenience tradeoff when it comes to things like E2EE and server-side search.