Most data-heavy sites treat "no value" as an edge case. It is not an edge case when you are publishing measurements — it is a value in its own right, and it needs a representation, a rendering rule, and a language.
Here is the specific problem. A page reports one number per row: mean duration, harmonics-to-noise ratio, a count of measured tokens. Some rows cannot be measured. The audio does not support it, or the measurement would be indistinguishable from noise.
You have four options, and three of them are wrong.
Option 1: fill in zero
Zero is a valid measurement. Reporting "0 ms" says the sound has no duration, which is a claim, not an absence. A reader cannot tell it apart from a real measurement of zero, and a downstream calculation will happily average it in.
This is the failure mode that looks most responsible — the table is full, nothing is blank, the page "works" — and it is the worst of the four, because the wrongness is invisible.
Option 2: leave the cell blank
An empty cell is ambiguous between "not measured" and "failed to load". If the rest of the table renders, a reader will assume a rendering bug; if several cells are blank, they will assume the page is broken.
Blank also invites the reader to fill it in themselves, from whatever they already believe.
Option 3: null in the data, dash on the page
This is the version that works, and it takes three layers to be honest about it.
Data layer. Distinguish "field absent" from "field present with no value". In practice: the key is missing when the concept does not apply (a consonant has no F0 row), and the key is present with an explicit null when the concept applies but could not be measured. Those are different facts and they must not collapse.
{ "symbol": "n", "duration_ms": 71, "hnr_db": null }
Rendering layer. A null renders as a dash, never as 0 and never as an empty cell — and the page carries one line explaining what a dash means, in the reader's interface language. The dash is a link to that explanation, not decoration.
Arithmetic layer. Any aggregate excludes nulls rather than treating them as zero, and reports n. "Mean 77 ms across 12 measured tokens" is honest; "mean 64 ms" computed over fifteen tokens where three were null is not.
The case that made the rule
The rule came out of one specific measurement that I decided not to publish: the duration difference between a word spelled with -ing and the same word spelled with -in'.
Measured across six pairs, the mean target-word length was 334 ms, range 262–383 ms — a useful table. But the pair that motivated the page, something against somethin', came out at 383 ms against 381 ms.
Two milliseconds is inside the noise of the aligner. Publishing "383 vs 381" invites a reader to conclude there is no difference, when the difference is real and lives in the transcription rather than the duration.
So that cell is null, the page shows a dash, and the note explains why: no duration is reported here; the difference between these two spellings is not a difference in length. The measurement itself still exists in the build output — it is just not published as a number, because a number would be read as an answer.
Why this is harder in two languages
The dash is interface, so it exists once per interface language. The absence is content, so it exists once per target language. The explanation for a missing cell therefore has to exist in every interface language and has to agree with the content — and it is easy to get one of the two right while the other drifts.
The version that holds up is unglamorous — the absence note lives in the same string table as every other caption, keyed by the reason for the absence rather than by the page. Three reasons today: not measurable from this audio, not applicable to this sound, and measured but not published.
The general rule
If a dataset has absences that a reader will notice, the absence needs to be part of the schema, not a rendering accident. Encode it, render it distinctly, exclude it from arithmetic, and say in the reader's language what it means.
A blank cell is a question. A dash with a reason is an answer.
The tables described here — including the null cells and their notes — are on phonemetrick.net.
Top comments (0)