Transcription is the first analytic act

A study returns with a recorder full of talk and a folder of video. Neither can be quoted yet. The moment someone writes the talk down, they make the first analytic decisions of the project.

This first of ten units argues that a transcript is already a theory of the data. What the transcriber preserves decides which claims the record can later carry. Those preserved features include overlap, pause length, intonation, gaze and gesture. Transcription is therefore an interpretive decision, not clerical preparation. The nine units that follow each take one step of the analysis.

The transcript is a theory, not a copy

The claim is old, and it is not a complaint about careless work. Ochs (1979) argued that transcription is not a neutral rendering of speech. It is a theory of speech. The transcriber decides what to represent, how to segment the stream, and which features count as meaningful. Those decisions pre-answer questions the analysis later treats as open. A transcript that breaks talk into sentences has decided that the sentence is the unit of interest. One that measures silence in tenths of a second has decided that the timing of silence matters.

Davidson (2009) reviewed three decades of transcription literature and found it presented as taken for granted in qualitative studies, naming the need for empirical studies of transcription in practice.

Notation is the lever on what a claim can carry

The clearest demonstration is the symbol set. Jefferson (2004) assembled the glossary of transcript symbols used in conversation analysis. The glossary carries overlap brackets, latching and gap notation, silence measured in tenths of a second, and marks for intonation, stress and volume. Each mark names the phenomenon it was invented to display. It lets a reader see a hesitation, an interruption or a rising pitch at all, and it is all a reader who cannot hear the recording has.

Hepburn and Bolden (2013) present the conversation-analytic approach as a method with a rationale. They cover the conventions, talk in other languages, visible conduct, and the concerns the approach has attracted. One objection they record is that the notation may encode the analyst's theory rather than present a neutral record. A system that is explicit about what it keeps has made its selectivity available for inspection. It has not abolished it.

Two defensible conventions make different documents from one recording

Interview research faces the same choice more sharply. Oliver, Serovich and Mason (2005) distinguish naturalised transcription from denaturalised transcription. Naturalised transcription captures every utterance in as much detail as possible. Denaturalised transcription corrects grammar, removes interview noise such as stutters and pauses, and standardises non-standard accents. The authors treat the decision as a powerful act of representation.

Their recommendation is the unit's practical rule: interrogate the transcription decisions and their possible impact on participants and findings. It does not ask for a neutral transcript, since there is none. It asks for one whose theory has been chosen and stated. The stakes reach past accuracy to representation.

What the written record cannot hold

Selectivity bites hardest once the data leave the audio channel. Ayass (2015) argues that the written version is a transformation of the data rather than its record. Working through examples from several authors, she asks how visuality, multimodality and artefacts can be captured at all, and finds the transcript caught between orality and literacy, and between artwork and figure.

Bezemer and Mavers (2011) ask in a social semiotic direction. How do researchers transcribe gesture and gaze, and how can a transcript show readers how those modes work alongside speech? They treat transcripts as artefacts whose meaning-making principles can be reconstructed. Their framework makes the loss measurable: a transcript that holds only the words has recorded one channel. Their own convention shows what that costs: gaze is written as "looks at" and a head movement as "shakes head". A claim about a reach for the mouse or a glance needs those features preserved, or it has no evidence.

Writing a speaker down encodes a judgement, and delegation moves the author

Bucholtz (2000) argues that transcription is political as well as theoretical. How a speaker is written, and whether their features are standardised or preserved, encodes judgements about the speaker and the language. She works the claim through two examples. In a police interrogation transcript, exchanges the recording carried are labelled "unintelligible" and mistranscribed, so the officer's words leave the record. In a newspaper transcript of a radio panel, a speaker of African American Vernacular English is written in standard orthography. The rule inserts the forms the language is assumed to omit, so "gonna" appears as "going to". Roberts (1997) makes the parallel argument for second-language data in TESOL.

Who holds the pen compounds the problem. Tilley (2003) argues that the transcriber's interpretive lens shapes the final text and can shape the analysis. Her case is Ken, a person hired to transcribe focus-group recordings whose difficulties with the work and use of voice-recognition software are reported in detail. When transcription is delegated, an interpretive act moves to someone whose assumptions the researcher has not examined, and the researcher still signs the finding.

Verbatim is an assumption that can be measured

All of this would be rhetorical if transcripts were accurate, and they are not, in specifics. Poland (1995) observes that the requirement that transcripts come close to a verbatim account is widely acknowledged. It functions more often as a tacit assumption than as an empirically verified assertion, and routine review of quality before analysis is barely discussed. He proposes ways to avoid, detect and remedy the threats to it, and concedes that error itself is socially constructed and open to correction. That concession is now partly dated, because Wollin-Giering and colleagues (2024) have measured what Poland treated as unmeasured, for nine automatic transcription tools.

MacLean, Meyer and Estable (2004) take the problem into practice. They examine the circumstances that bring about a consistent set of troubles: voice-recognition systems, notation choices, processing against touch typing, the effect of the transcriptionist, emotionally loaded recordings, and cultural and second-language differences. Give the transcriber a notation system with worked examples. Spot-check transcripts as they develop rather than only at the end. Lapadat and Lindsay (1999) give the cross-disciplinary summary: transcription is theory laden, and the choices researchers make constrain the interpretations they can draw.

Contested question: can a transcript be neutral?

The literature agrees that transcription is consequential. It disagrees about whether the consequence can be managed. The disagreement is durable because both sides accept the same facts and differ over the inferences those facts support.

The first position holds that a defensible record is achievable, so the problem is technical and the answer is disclosure. Its evidence is mostly prescriptive rather than comparative. Poland (1995) argues it in a methods article and reports no sample from which an error rate could be estimated, though Wollin-Giering and colleagues (2024) have since measured accuracy across nine tools. MacLean, Meyer and Estable (2004) name the troubles that recur and give remedies drawn from practice, and Hepburn and Bolden (2013) set out the conventions in a handbook chapter. The side shows that errors can be found, and a procedure for finding them. It has not shown that two defensible conventions leave a claim unchanged.

The second position holds that notation necessarily encodes the analyst's theory, so disclosure only makes the theory visible. Its evidence is conceptual and case-based. Ochs (1979) states the argument in a chapter with no data. Lapadat and Lindsay (1999) sharpen it in a cross-disciplinary review of the literature. Bucholtz (2000) works it through two examples, a police interrogation and a newspaper panel. Ayass (2015) argues it theoretically. Tilley (2003) carries it in a single case, the transcriptionist she calls Ken. What would settle it is a controlled comparison, and neither side has run one; Davidson (2009) names that absence.

The strongest counter-argument to this unit's own claim is that it overstates the case. Where a convention states in advance what it keeps, a transcript can be checked against the recording on exactly those features, and the check is measurement rather than theory. On that reading a transcript is a fallible instrument with a known sensitivity, not a theory of the data. The claim holds only for what the convention never declared.

What to do while the dispute is open

Four moves keep a study defensible meanwhile. State the convention and what it keeps. Keep the recording, so the transcript stays auditable. Transcribe a subsample twice, and report where the claims diverge. When a decision rests on a transcript feature, name it and say what would change had the notation kept something else.

Where the argument is heading

Ayass (2015) argues that the evolution of recording media altered the status of transcripts. Bezemer and Mavers (2011) have begun the answer by treating multimodal transcripts as artefacts with reconstructable principles. Nikander (2008) shows that translation adds a further layer. Underneath runs a change in who does the work. Tilley (2003) and MacLean, Meyer and Estable (2004) record voice-recognition software entering the transcriber's workflow; Wollin-Giering and colleagues (2024) measure how it performs there. The emerging demand is not for a neutral transcript but for an account of how the rendering was produced.

The sources this unit rests on

• Ayass, R. (2015). Ayass argues that the transcript is a transformation of the data rather than its record, and that the medium shift left it caught between orality and literacy. Her discussion of what multimodality and artefacts lose gives the unit its account of the speech-only record's limits.
• Bezemer, J. and Mavers, D. (2011). Bezemer and Mavers ask how researchers transcribe gesture and gaze and how a transcript can show readers how those modes work alongside speech, treating transcripts as artefacts whose meaning-making principles can be reconstructed. Their framework is the unit's clearest case that a speech-only transcript has recorded one channel rather than the interaction.
• Bucholtz, M. (2000). Bucholtz argues that transcription is political as well as theoretical, showing through examples that how a speaker is written, whether features are standardised or preserved, encodes judgements about the speaker and the language. She supplies the unit's argument that the stakes reach beyond accuracy to representation: the abstract states that transcription involves both interpretive decisions about what is transcribed and representational decisions about how.
• Davidson, C. (2009). Davidson reviews three decades of the transcription literature, between 1979 and 2009, and finds it central to qualitative research yet presented as taken for granted. She calls for more empirical studies that examine transcription in practice, which is the unit's evidence that the question is unresolved.
• Hepburn, A. and Bolden, G. B. (2013). Hepburn and Bolden set out the conversation-analytic approach as a method with a rationale, covering the conventions, talk in other languages, visible conduct and the concerns the approach has attracted. The section on those concerns supplies the unit's objection that the notation may encode the analyst's theory rather than a neutral record.
• Jefferson, G. (2004). Jefferson's glossary sets out the transcript symbols of conversation analysis, including overlap brackets, latching and gap notation, silence measured in tenths of a second, and marks for intonation, stress and volume, each introduced with the phenomenon it displays. It is the concrete apparatus behind the unit's claim that notation is selective.
• Lapadat, J. C. and Lindsay, A. C. (1999). Lapadat and Lindsay present a cross-disciplinary review and conclude that transcription is theory laden, so the choices researchers make enact the theories they hold and constrain the interpretations they can draw. That is the unit's thesis in its sharpest published form.
• MacLean, L. M., Meyer, M. and Estable, A. (2004). MacLean, Meyer and Estable examine the circumstances that produce recurring transcription problems, among them voice-recognition systems, notation choices, processing against touch typing, the effect of the transcriptionist, emotionally loaded recordings, and cultural and second-language differences. They recommend practices for working with transcriptionists and checking the result.
• Nikander, P. (2008). Nikander argues that transcripts are produced for particular analytic purposes and so range from broad verbatim accounts to refined interactional renderings, and that translation compounds the transformation. The point the unit draws is that the level of detail follows from the question rather than from a fixed standard.
• Ochs, E. (1979). Ochs argues that transcription is not a neutral rendering of speech but a theory of it, because the transcriber decides what to represent, how to segment it and which features to treat as meaningful, and those decisions pre-answer questions the analysis later treats as open. The chapter is the origin of the unit's governing claim.
• Oliver, D. G., Serovich, J. M. and Mason, T. L. (2005). Oliver, Serovich and Mason distinguish naturalised transcription, which captures every utterance in as much detail as possible, from denaturalised transcription, which corrects grammar, removes interview noise such as stutters and pauses, and standardises non-standard accents. They treat the decision as a powerful act of representation and recommend interrogating it as part of the design.
• Poland, B. D. (1995). Poland argues that the requirement that transcripts approximate a verbatim account is a tacit assumption rather than an empirically verified assertion, and that routine review of transcript quality before analysis is barely discussed. He reviews the threats to quality and proposes ways to remedy them, conceding that error is socially constructed.
• Roberts, C. (1997). Roberts writes about the politics of transcription in TESOL, arguing that decisions about how to represent spoken language, especially second-language speech, carry assumptions about the speakers and the setting. The commentary carries the unit's argument into applied language research, where the represented speech is not the analyst's own.
• Tilley, S. A. (2003). Tilley interrogates the practice of hiring someone other than the researcher to transcribe, arguing that the transcriber's interpretive lens shapes the final text and can shape the analysis. Her case is Ken, whose difficulties with the work and use of voice-recognition software are reported in detail.
• Wollin-Giering, S., Hoffmann, M., Hofting, J. and Ventzke, C. (2024). Wollin-Giering and colleagues compare nine automatic transcription tools on an English and a German interview, measuring word accuracy near 85 per cent for most services, 93 per cent for Whisper, and about 50 per cent for the German interview. Every transcript stops between verbatim and gisted, so manual review remains necessary, and the measurement dates Poland (1995) by supplying the accuracy he treated as unmeasured.