Transcription conventions decide your model's ceiling

Speech projects tend to treat transcription as a solved input: get the words down accurately and move on to the modelling. But "the words" is doing a lot of work in that sentence, and the decisions hiding inside it shape everything downstream.

The questions nobody asks early enough

• Is it "twenty twenty six", "2026", or both depending on context?
• Do you transcribe "um" and "uh", or remove them?
• When two people talk at once, is that one segment or two?
• Does a false start get written, or silently corrected?
• How is an unintelligible passage marked — and is it marked at all?

Each has a reasonable answer. None has a universally correct one. What matters is that the answer is decided once, written down, and applied to every hour of the corpus.

Why the cost of changing is so high

Change a convention halfway through and you do not have one dataset with a minor inconsistency. You have two datasets that disagree, and a model that will learn the disagreement as though it were signal. The only real remedy is re-transcribing everything produced under the old convention.

This is why conventions belong in the specification phase alongside the model design, not in an onboarding document written after transcription has begun.

Timing is half the deliverable

An accurate transcript with loose timestamps is one of the most common reasons a technically correct dataset trains poorly. Alignment between audio and text is what lets a model associate sound with meaning; if segment boundaries wander by a few hundred milliseconds, that association degrades in ways that are difficult to diagnose later.

Decide the granularity your model consumes — segment, utterance or word — and specify the tolerance. "Roughly aligned" is not a specification.

Clean audio is the wrong target

There is a strong pull toward collecting and keeping only clear recordings. Background noise, crosstalk and heavy accents are harder to transcribe, produce more disagreement, and make quality metrics look worse.

The better approach is to keep the difficult material and be explicit about how it is handled: mark unintelligible spans rather than guessing, label overlap rather than discarding it, and let the model see the conditions it will meet.

Verbatim or clean read

This is the single largest fork, and it should be decided by what the model is for. A system that has to handle spontaneous speech needs the disfluencies, repetitions and self-corrections that a clean-read transcript removes. A system summarising meetings mostly does not.

The failure is not choosing wrong. It is choosing implicitly — letting each transcriber decide how much to tidy, which produces a corpus that is neither and teaches the model that hesitation is sometimes worth modelling and sometimes not.

Speaker labels have to be stable

Within a single recording, speaker one must be the same person from the first minute to the last. Across a corpus, whether speaker one means anything at all is a separate decision — usually it should not, unless you are deliberately building speaker-linked data.

Getting this wrong is easy and quiet. Labels reset per file, a transcriber swaps them halfway through a long session, and a diarisation model learns from identities that do not hold.

Specify these before recording

• Verbatim or clean read, with examples of each on the same passage
• Number, date and currency formatting
• How overlapping speech is segmented and attributed
• The marker for unintelligible audio, and how much context is required before guessing
• Timestamp granularity and the tolerance you will accept