What is Transcription?

Transcription looks simple and is in practice where speech data projects go wrong most often. The same recording transcribed by pronunciation and by standard spelling produces two entirely different texts.

That is why any serious project writes a transcription guideline first, fixing how punctuation, particles, loanwords and spoken numbers are handled.

Levels of transcription detail

Transcription is not one product. The first choice is verbatim or cleaned. Verbatim writes what was said, including the false starts, the repeated words and the filled pauses; cleaned removes them and produces text that reads like prose. Neither is better in general. Speech recognition training usually wants verbatim, because the model has to handle disfluent speech. Subtitling and search usually want cleaned text.

The second choice is whether the text is tied to time. Plain transcription gives one text file per recording. Aligned transcription gives segment or word boundaries, so every word maps to a position in the audio. Alignment is additional work, and it is what makes a transcript usable for subtitling, for searching inside a recording, and for training a streaming system.

The third is what gets marked beyond the words: whether laughter, coughs and background speech are noted, whether speaker turns are marked, and whether the text follows standard spelling or records the sounds as produced, so that a word like gonna is written as it was said. Each is a decision the guideline has to make, and each changes what the delivered text is good for.

What to specify when buying

Ask to see the transcription guideline before you ask for a price. A workable one fixes the spelling standard and which variety of it, how numbers, dates and currency are written, whether punctuation and capitalisation are added, how filler words and partial words are handled, how loanwords and proper nouns are spelled, and the item most often missing, how unintelligible audio is marked.

The audio side matters as much. Sample rate, channel layout, how files are segmented, and how each audio file is named to match its transcript. Ask what proportion of the work is double-checked by a second transcriber, and what disagreement rate the supplier measures on that check. A supplier who can quote that figure is running a process.

Left vague, transcription goes wrong quietly. A supplier who cleans the text without saying so makes the data unusable for training a system that must transcribe disfluent speech, and the buyer usually finds out after delivery. Hours delivered without alignment cannot be used for anything that needs timing. Because transcription is priced per audio hour, the hour count and the rate are the numbers the quote turns on.