Soundex may not be perfect, but it seems like a reasonable approach, at least for your specific example of English/Hebrew matching.
You definitely can't use the rule about preserving the first letter of the name, but I never liked that even for the Latin alphabet (because I'd have to look under both "E" and "Y" for my mother's family name). I recommend just treating the first letter like all the others.
Then it's just a matter of mapping the Hebrew letters to Soundex codes. You don't really need an intermediate English transliteration; just code the Hebrew → Soundex mapping directly.
- בוףפ → 1
- גזחךכסקש → 2
- דטת → 3
- ץצ → 32
- ל → 4
- םמןנ → 5
- ר → 6
- אהיע → ignored
However, because Soundex is English-centric, it may not correctly handle certain ambiguities in the Hebrew pronunciation:
- ו is mapped to 1 (like English V) in the list above, but it often represents O, U, or W, in which case it should be ignored in Soundex.
- ח is hard to classify due to its lack of an English equivalent. I put it in category 2 because this (1) matches the "ch" transliteration, and (2) allows ך/כ to have the same category with or without a dagesh.
- Ashkenazi pronuncation would split ת between categories 2 and 3.
To deal with this, you could generate multiple Soundex keys for a string. E.g., "שבת" would map to both 212 and 213.
Similar mappings can be made for Greek:
- ΒΠΦ → 1
- Ψ → 12
- ΓΖΚΞΣΧ → 2
- ΔΘΤ → 3
- Λ → 4
- ΜΝ → 5
- Ρ → 6
- ΑΕΗΙΟΥΩ → ignored
or Russian:
- БВПФ → 1
- ГЖЗКСХЧШЩ → 2
- ДТ → 3
- Ц → 32
- Л → 4
- МН → 5
- Р → 6
- АЕЁИЙОУЪЫЬЭЮЯ → ignored
(Note that some of the 2's might be 32's, depending on your transliteration convention.)
A similarity "rating" can be obtained based on a metric like longest common subsequence length or Levenshtein distance on the Soundex values.
For example, you can define the "similarity" between two strings as 2*lcslen(A, B)/(len(A)+len(B)) to obtain a score between 0 and 1.
