They pad a short digit sequence to the conventional three-digit portion; they are not extra sounds.

Can differently spelled names share a code?

Yes. That is an intended feature, although unrelated names can also collide.

Does the initial letter matter?

Yes. The conventional code retains it, so some pronunciation-related variants with different initial letters can remain separated.

AI Glossary / Retrieval and knowledge systems / Intermediate

Soundex

Definition

Soundex is a phonetic indexing method that groups selected spelling variants under a short code. In the conventional surname system described by the US National Archives, the code contains the surname's initial letter followed by three digits. Consonants are grouped into sound-related classes, while other letters and repeated codes are handled by specific rules. Soundex helps find candidate records; it is not a unique identifier or a complete pronunciation model.

What it means

Exact spelling can be unreliable in older records, manual data entry, or searches made from memory. A phonetic index offers another route to a possible record when the spelling differs. The same compression that supports this use also creates collisions between unrelated names.
The familiar letter-and-digits system is only one member of a broader family of phonetic methods. Daitch-Mokotoff Soundex, for example, has a different format and rules. Storing both under an undifferentiated field called soundex can create confusion unless the method is recorded.

How it works

Retain the initial letter and encode later consonants by groups: B/F/P/V use 1; C/G/J/K/Q/S/X/Z use 2; D/T use 3; L uses 4; M/N use 5; and R uses 6. Keep three digits and pad with zeros when needed.
Repeated adjacent codes collapse, including interactions with the initial letter's class. Vowels separate same-code consonants, while H and W do not in the National Archives rules. The order of these operations matters. Simply deleting every ignored letter before collapsing duplicates can produce an incorrect result. Use one documented convention consistently for indexing and queries.

Three practical examples

A family-history search can locate possible surname spellings in an index before inspecting the underlying records.
A directory can offer additional name candidates when an exact query returns no useful result.
A data-integration test can use Soundex agreement as a broad screening feature and evaluate the unrelated pairs it introduces.

Think of it this way

Think of filing names in drawers labeled by a rough sound-based abbreviation. A spelling variation may lead to the same drawer, making the record easier to discover. However, several unrelated people can be filed there. The analogy is limited because the algorithm follows spelling rules rather than hearing or understanding how each person actually pronounces a name.

Worked example

Consider an original manual trace for Brown, Browne, and Braun under the National Archives convention. Brown retains B. R contributes 6, O and W contribute no digits, and N contributes 5. Padding the two-digit sequence 65 with a zero gives B650. Browne follows the same relevant steps, with the final E adding nothing, so it also yields B650. Braun retains B, contributes 6 for R, skips A and U, and contributes 5 for N, again giving B650.
These three spellings therefore share a retrieval key. This is a statement about the rule output, not a claim that every person with these names has the same pronunciation or identity.
To isolate the separator rule, examine two explicitly invented letter strings rather than real surnames: BABA and BHBA. For BABA, the A between the two B letters separates their code classes. The initial B is retained as a letter, and the later B contributes 1. Padding gives B100. For BHBA, H does not separate the repeated B class. The later B contributes no additional digit, so the result is B000. This demonstrates why deleting ignored letters first would lose information needed to apply the duplicate rule correctly.
Now suppose a fictional index contains eight records under B650, of which two are relevant to a particular search. The key lookup returns all eight, giving candidate precision 2/8, or 25 percent. If both relevant records are present, recall within that defined target set is one. A shared code has recovered candidates but has not ranked them or established identity. The search still needs original record details and contextual evidence. These traces and counts were checked manually; they are not a claim of running a library or measuring a real archive.

Illustrative case study

Fictional case: a local-history team imports a surname index into a new search system. Most examples work, but names containing H or W between same-class consonants produce inconsistent keys.
The team compares the implementations with the National Archives rules and adds explicit compatibility examples. It rebuilds the affected derived index from original spellings. Search results continue to display the source records, and researchers can still use exact spelling and other search fields when phonetic matching is unhelpful.

Common misconceptions and mistakes

Do not compare Soundex codes as ordinary numbers. The digits are class labels, and subtracting codes does not produce a meaningful pronunciation distance. Do not infer relationships, ancestry, or other personal characteristics from a collision.
The method is not a universal multilingual matcher. It can miss useful variants and group unrelated names, particularly when the collection differs from the spelling conventions the rules handle well. A code also cannot correct every transcription problem. Missing names should remain missing rather than being assigned a valid-looking key that creates false agreement.

Business use

Soundex can be a simple candidate-generation option when exact spelling is too restrictive. Its value depends on the number of relevant variants it recovers relative to the additional review burden.
Record which Soundex convention is used, preserve original names, and verify that indexing and querying normalize input consistently. Test difficult nonmatches as well as obvious variants. In record linkage, combine code agreement with appropriate independent evidence. Measure candidate recall separately from the accuracy of final linkage decisions, because retrieving a candidate and deciding that two records belong together are different tasks.

Check your understanding

Why do the toy strings BABA and BHBA receive different codes under the stated rules?

Answer: A separates the repeated B code class, allowing the later B to contribute 1. H does not separate it, so the later B is suppressed.

Related terms

Learn next

Apply the separator rule to the toy strings BOB and BWB, keeping the initial letter convention explicit. Then explain why a search interface should show original spellings and record context even when all returned candidates share the same code.

Sources

Primary operational source: US National Archives, Soundex System: https://www.archives.gov/research/census/soundex . Brown/Browne/Braun traces, invented separator strings, and candidate-set arithmetic are original exercises following its published rules. This entry describes that conventional surname format, not every algorithm carrying the Soundex name.

Previous: Sort Operation Next: Source-filter model