Conversation Analysis: the machinery of talk-in-interaction

The unit is the sequence, not the sentence. Utterances are actions positioned in sequences — a first pair part (question, invitation) makes a second (answer, acceptance/declination) relevant, and whatever comes next will be heard against that relevance. Silence after an invitation is not nothing; it is a noticeably absent acceptance, and speakers treat it as such.

The next-turn proof procedure is the method’s epistemology: how a recipient responds displays how they understood the prior turn, so analysts’ claims about what an utterance was doing are checked against what participants themselves showed they took it to be doing.

Deviant case analysis replaces statistics: a candidate rule (say, that dispreferred responses are delayed and mitigated) is tested by hunting cases that appear to violate it — and the analysis succeeds when the violation turns out to be oriented to by participants, confirming the rule’s reality for them.

The founding paper is Sacks, Schegloff and Jefferson’s A simplest systematics for the organization of turn-taking for conversation — the turn-taking model whose rules (one speaker at a time as a managed achievement, transition-relevance places, allocation techniques) remain the field’s spine.

Gail Jefferson’s transcription system is the method’s second foundation: a notation capturing overlap, latching, pauses to the tenth of a second, prosody, and breathiness. Fidelity at this level is not fussiness — CA’s findings live exactly in the details other methods clean away.

The comprehensive modern reference is Sidnell and Stivers’ Handbook of Conversation Analysis (Wiley), which covers the organisations — turn-taking, sequence, repair, word selection — and the institutional-talk programme applying them to medicine, courts, classrooms, and calls.

• Choose it where the interaction is the outcome: consultations, helplines, interrogations, sales calls — settings where what happens is built turn by turn and small design choices have measurable consequences.
• Choose it when you have (or can get) recordings. No recordings, no CA: field notes and recollections cannot support claims at the required grain.
• Accept the context discipline: gender, power, institution enter the analysis only as participants observably orient to them. If your question requires importing social structure, CDA or ethnography fits better.
• Do not choose it for what people think or experience — CA is resolutely about observable conduct, and interviews about talk are, for it, just more talk.

A team analysed 118 recorded primary-care consultations to study treatment recommendations. The collection assembled every recommendation sequence — 174 instances — transcribed in Jefferson detail.

The machinery told a precise story. Acceptances came fast and unmarked (“okay, yeah”); resistance almost never arrived as refusal. Instead patients withheld acceptance — a 0.4-second silence, a “mm”, a repeat of the drug’s name with rising intonation — and physicians demonstrably oriented to these micro-signals, responding with justifications, dosage softenings, or alternatives before any explicit objection existed. The next-turn proof was doing the analytic work: the physicians’ conduct showed that withheld acceptance was being heard as resistance.

Deviant cases sharpened the finding: in the handful of consultations where physicians did not respond to withholding and simply proceeded, prescriptions were issued — and follow-up records showed those were disproportionately the prescriptions never filled. The applied conclusion — that unaddressed micro-resistance predicts non-adherence, and that pausing after recommendations invites it into the open — has since informed communication training, a straight line from tenth-of-a-second transcription to clinical practice.

• Working from clean transcripts. Standard orthographic transcription deletes the phenomena; CA on cleaned data is analysis of an artefact.
• Single-case overreach. One elegant analysis without a collection is a noticing, not a finding.
• Importing context. Explaining a turn by the speaker’s gender or status, without showing participants orienting to it, breaks the method’s evidentiary rule.
• Treating CA terms as labels. Tagging turns “repair” or “dispreferred” without analysing the sequential work is classification wearing analysis’s clothes.
• Ignoring the body on video data. Gaze and gesture are part of turn design; audio-only claims from video corpora should say so.
• Skipping deviant cases. The collection’s exceptions are the test; unexamined, the rule is an impression.

CA’s precision is bought with narrowness: it speaks about observable interactional conduct and is designedly silent about experience, motive, and structure. Questions about why people talk as they do, biographically or socially, exceed the warrant.

The context discipline, its greatest strength, is also contested — critics argue that refusing exogenous context blinds the method to power that participants cannot safely display. The institutional-talk programme narrows but does not settle the argument.

And the costs are stark: Jefferson transcription runs to an hour or more per minute of talk at full detail, collections take corpora, and the skill floor is high. CA repays the investment only where interactional detail is genuinely where the answers live.

Discourse analysis shares the focus on language-in-use but admits interpretive resources CA refuses.

Critical discourse analysis imports exactly the contextual apparatus CA brackets — the two methods define each other’s boundaries.

Narrative analysis attends to storytelling structure at a coarser grain than sequential organisation.

In-depth interviews generate accounts; CA would study the interview itself as interaction.

CA’s bottlenecks are transcription and collection-building. Automatic transcription with speaker diarization and timestamps — Evidano’s transcription does both — produces the accurate first pass onto which Jefferson detail is layered for target segments, cutting the hours where they are least analytic. Coding every candidate instance of a practice across a corpus, with each tagged excerpt linked to its exact location and audio, turns collection-building from spreadsheet archaeology into retrieval.

The detail passes and the sequential analysis are the method itself and stay handmade: no automatic transcript hears a 0.2-second silence as an action, and no model should be trusted to.