Discourse analysis transcript: From recording to counting words

Discourse, thematic and content analysis compared

The boundary between thematic and content analysis is a common question. Roughly, content analysis can turn categories into numbers, while thematic analysis cares less about how often a pattern occurs than what it says about the research question. Vaismoradi and colleagues discuss that boundary, and Braun and Clarke’s guide is a good start. For critical discourse analysis, van Dijk’s principles are widely cited.

How much detail should the transcript carry?

The method sets the detail. Thematic analysis may do with cleaned text. Discourse analysis often needs a verbatim one. Fine-grained work like conversation analysis also records pause length, overlap and intonation, usually in Jefferson notation.

Automatic text does not give you that detail. LiveType does not write hesitation sounds like “um”, and the model sometimes skips words like “you know” as well. If they matter, add them by ear using transcription conventions before you count. Counting an unfinished text gives an unfinished answer.

Counting a phrase by hand

Ctrl+F in Word or a browser is the fastest way, and a plain text export can go into counting tools. Set these rules first:

• Do you count each speaker apart? Speaker names matter here.
• Do the interviewer’s words count?
• Is one person’s immediate repetition one or two?

LiveType does not count words or draw word clouds. You do that with the text in hand.

What a search misses: Word forms and errors

A plain search for one word finds too little or too much:

So list the forms in the text first, group the relevant ones by hand, and then count. Languages with many suffixes, such as Turkish, make this harder still, since one root takes dozens of forms.

Transcription errors change counts as well. If the model wrote a word as a similar-sounding one, that occurrence is not on your list. Compare the sections you take figures from with the audio.

What a word cloud shows and hides

A word cloud draws the most frequent words large. It is useful as a first look and in a slide, but it is not analysis:

• It shows: Roughly which words stand out.
• It hides: Whether a word was used for or against something, who said it, and what was never said.
• It misleads: Tools that do not merge word forms spread the same word across several entries.

Before making a cloud, remove speaker names and function words from the plain text file, or they will be the largest words.

Counting with an LLM, and checking it

Large language models are quick at rough sorting but unreliable at counting: They may make the number up. Ask for evidence instead of a figure:

In the text below, list every sentence that contains any form of
the word "migrant", exactly as written. Do not change, summarize
or guess. At the end, give the count for each form.

Then run three checks:

1. Find the list in the text. If a sentence is not in the text, throw the result away.
2. Test a small part. Compare with a section you counted by hand to see how far off it is.
3. Keep identities out. Anonymise personal data before giving it to an outside service.

Where transcription stops and analysis starts

LiveType stops at the transcript: It is not analysis software, assigns no codes and suggests no themes. What it does is turn the talk into text quickly without losing the link to the sound. Being able to return to the sound from a paragraph’s time makes it easier to check a quote’s context. If your source is an interview, read qualitative interview transcription. If it is a public statement, read press conference coverage. A word’s frequency alone is not discourse analysis: The reading is yours.