Archive intelligence / Research guide

Podcast dialogue-act datasets: conversational function beyond speaker labels

Specify podcast dialogue acts, repairs, backchannels and clarification with context, uncertain labels and auditable annotation. Includes authored examples.

FIELD NOTESBY FIUND
01Session02Edition03Archive

Recordings · references · lineage

Cited references
5 cited sources
Inside this guide
1 worked example
Reading time
8 minutes
Reviewed
2026-11-13

Prepared by the Fiund research desk from the cited public documentation. Technical examples describe their evidence and limits. Research method & editorial contact · Suggest a correction

Reading tools
The takeaway

A podcast dialogue-act dataset labels what a contribution does in its conversational context: asking for clarification, acknowledging a speaker, repairing a statement or giving an answer. Useful references need a segmentation convention, surrounding turns, clear label definitions and a review process for ambiguous cases. Speaker labels and transcripts alone do not establish these functions.

Core unit
Contextual functional segment or linked exchange
Separate layers
Words, speaker timing, audio events and dialogue function
Quality rule
Retain disagreement and unresolved labels
Example
Six fictional authored turns; no real audio

Your reading path

Start with the right questions.

  1. 01Choose the interaction problem before choosing a label inventory
  2. 02Use primary annotation research without pretending meetings are podcasts
  3. 03Distinguish audible events from interpreted conversational functions

From principle to practice

Distinguish audible events from interpreted conversational functions

LayerExample fieldWhat the field does not establish
Open the worked example

Choose the interaction problem before choosing a label inventory

A turn-taking model, interview-search system and conversational assistant may use the same recording but need different references. A search system may care about questions and their answers. A dialogue agent may care about acknowledgments that should not trigger an interruption. An editing tool may care about self-correction and whether a removed span changes a proposition. Define the intended decision first, then collect labels that make that decision observable.

Diarization answers who spoke when; transcription records words; dialogue acts describe communicative function in context. These layers can disagree without either being automatically erroneous. A short utterance can be an answer, acknowledgment or request to repeat depending on surrounding turns. A diarization boundary should therefore be treated as an input to semantic segmentation, rather than an unquestioned dialogue-act boundary.

Use primary annotation research without pretending meetings are podcasts

The AMI Corpus annotation layers and reliability documents multiple annotation layers, annotation coverage and reliability information. Its AMI Corpus annotation guidelines separately cautions that the existence of guidelines does not establish that corresponding annotations are available. This is a useful procurement lesson: obtain the actual layer, version, coverage and export, rather than assuming that a paper or manual implies a complete downloadable label set.

AMI is a meeting corpus, not a representative sample of podcast interviews. Its annotation practice can inform a rubric, but adapting it requires reviewing conversational style, episode structure, editing and speaker roles. A rehearsed monologue, two-host discussion and adversarial interview have different opportunities for clarification, floor management and repair. Report those differences instead of calling every conversational source interchangeable.

Distinguish audible events from interpreted conversational functions

The PodcastFillers dataset annotation vocabulary describes manually annotated audio events alongside automated transcript output; those manually marked events are not simply present in its automatic transcripts. This supports separating an acoustic event layer from the words and from a dialogue-act layer. A filler occurrence is not, by itself, a gold label for hesitation, uncertainty or an intention to keep speaking.

For example, “like” can be lexical content or a discourse marker. A brief “mm-hm” can acknowledge continued listening or agree with a proposition. Laughter may occur during an answer without replacing the answer function. Keep the observable token/event and the proposed interpretation in separate fields. Do not map every filler, laugh or overlapping sound to a semantic act through a fixed lookup table.

Compare the evidence
LayerExample fieldWhat the field does not establish
Transcriptutterance_textCorrect conversational function
Audio eventfiller/laughter intervalSpeaker intention
Interaction timingoverlap intervalAn intentional interruption
Dialogue actclarification_questionA universal label across all taxonomies

Define a unit that can survive multi-function speech

An entire speaker turn may contain a question, a repair and a qualification. Decide whether the annotation unit is a turn, a functional segment inside a turn or a linked collection of segments. If multiple labels are allowed, define whether they are simultaneous functions, ordered subsegments or alternatives under uncertainty. Otherwise a multi-label export becomes impossible to interpret consistently.

Pin transcript revision and segmentation version. Store the source turn identifier, segment character offsets and, where real audio is reviewed, the corresponding time interval and clock convention. Preserve the full unedited turn as context. When an editor changes punctuation or combines transcript segments, a character span can move; stable identifiers and revision hashes make that change detectable.

Inspect the authored example before requesting recorded data

The dialogue-act JSONL example contains six completely authored turns between fictional speakers. The guest corrects a session count; the host acknowledges and later asks which task makes incomplete sessions useful. There is no underlying recording, no real participant, no acoustic measurement and no claim that an annotator adjudicated these labels.

The proposed labels illustrate separate functions: confirmation question, self-repair, informative statement, clarification question and clarification answer. The “mm-hm” record deliberately preserves a competing agreement interpretation. Audio fields are null and audio_available is false. This makes the example useful for discussing a schema while preventing a text-only mock from being misrepresented as training-ready acoustic evidence.

The companion adjudication example records that unresolved decision and the missing evidence. It is a review protocol illustration, not a completed quality report. An importer should retain this status and reject conversion of unresolved alternatives into a single unquestioned reference label.

Attach questions, answers and repairs to their context

A clarification answer should point to the clarification question it resolves. A self-repair should identify the reparandum and the repair, if that distinction matters for the task. A response can refer to an earlier turn rather than the immediately preceding speaker contribution. Maintain relation identifiers independently of simple adjacency, and document how interruptions or inserted acknowledgments affect those relations.

Context windows should be explicit: previous two turns, a complete question-answer exchange or a longer topic segment are different annotation conditions. Reviewers deprived of the relevant earlier proposition may reasonably choose different labels. Keep the context presented to annotators in the provenance record, along with whether they could listen to audio. Model evaluation should disclose the available context rather than attributing all disagreement to model quality.

Do not infer interruptions or emotion from overlap alone

Overlapping speech is a timing observation. An interruption implies an interaction interpretation and may require prosody, the trajectory of the previous turn and what happens afterward. Cooperative completion, laughter and backchanneling can overlap without functioning as a hostile floor seizure. If interruption behavior matters, specify observable criteria and allow unresolved cases rather than assigning intent from an overlap flag.

Likewise, a correction does not prove embarrassment, and a filler does not prove cognitive difficulty. Keep affect or intention labels outside the scope unless the project has a defensible separate annotation protocol. For text-only samples, explicitly exclude acoustic claims. For recorded samples, retain source channels and timing evidence sufficient for authorized reviewers to revisit the judgment.

Make disagreement useful rather than hiding it

Write definitions, inclusion/exclusion examples and borderline cases for each label. Have independent reviewers annotate a shared calibration sample before expanding the assignment. Preserve their original judgments, the adjudicated decision and the reason for any override. A changed rubric needs a version and a decision about whether earlier labels require review; silently mixing versions undermines a clean benchmark.

Report disagreement by class and source condition rather than only one aggregate agreement score. A frequent informative-statement class can dominate a summary while rare clarification or repair labels remain unreliable. Missing audio, poor transcription, ambiguous segmentation and genuine alternative interpretations are different causes. Track them separately so a buyer can decide whether to obtain better evidence, change the taxonomy or exclude a case.

Build an evaluation split around sources and repeated interactions

Keep the same episode, derivative clip and corrected edition within a single partition. Consider grouping by show, host or recurring guest when the intended evaluation is generalization beyond familiar speaking styles. Randomly distributing neighboring turns across train and test can expose the same conversation context on both sides. Document the actual grouping constraint and any exceptions.

Evaluate boundary quality separately from label classification where both are learned. For labels, report per-class precision and recall, support counts and an explicit policy for multi-label or unresolved references. Include negative cases, such as an ordinary lexical “like” and an answer without an explicit question marker. This page supplies no measured scores; a proposed metric becomes evidence only after a defined evaluation is run.

Scope access and work units before discussing price

The PodcastFillers audio and metadata license notes distinguish audio and metadata terms and include research/commercial-use restrictions for the published artifacts. Those source-specific terms do not establish a license for a custom Fiund collection, and public accessibility does not settle the intended training use. Review the exact selected material and artifact terms before including it in a commercial plan.

Separate source access, channel preparation, transcription, segmentation, acoustic-event marking, semantic-act labeling, independent review and adjudication. A hypothetical one-hundred-turn pilot with two independent passes requires two hundred initial turn judgments before adjudication; it does not guarantee one hundred accepted single-label examples. This is an effort-unit illustration, not a market rate or a claim about Fiund throughput.

Quote the accepted unit explicitly: reviewed turn, functional segment, linked exchange or accepted audio hour. Repeated labels, rejected sources and unresolved units should remain visible in the accounting. A short difficult exchange can require more review than a long simple monologue, so episode duration alone is an incomplete comparison between proposals.

What to bring to a Fiund dialogue-data inquiry

Send the target interaction task, desired source styles, language and speaker requirements, need for actual multichannel audio, label convention, context window, output schema and acceptance rules. Include a small non-sensitive example of the desired annotation, and identify whether it is authored or sourced. Specify intended use so source access and preparation can be reviewed together.

Contact [email protected] to discuss a podcast dialogue-data requirement. Source availability, applicable terms and preparation feasibility are project-specific; these authored artifacts are not inventory. For the underlying source layers, read the diarization and time-aligned transcript guides, and use the question-answering guide when the primary target is supported answers rather than conversational function.

Related sourcing decisions

Raw multitrack podcast datasets: participant tracks, synchronization and rights — Raw multitrack podcast data explained: participant recordings, timing drift, source manifests, contributor scope and accepted session-hour budgets.

Time-aligned podcast transcript datasets — Podcast transcripts aligned to the correct audio version: word and turn timing, source manifests, review levels, provider comparisons and transparent pricing context.

Questions buyers ask

Can diarization labels be converted directly into dialogue acts?

No. They identify speaker activity. Conversational function needs context and a separate label convention; one diarized turn may contain several functions.

Is overlapping speech always an interruption?

No. Overlap can include backchannels, laughter and cooperative completions. An interruption judgment requires a defined interaction protocol and suitable evidence.

Does the sample contain real podcast voices?

No. The example is original authored text with fictional speakers, null audio timings and explicitly proposed labels. It is not a captured or reviewed podcast dataset.

Are AMI and PodcastFillers commercial Fiund inventory?

No. They are external primary references for annotation design. Their availability and source-specific terms must be evaluated separately from any custom project.

Sources & research notes

Public documentation reviewed 2026-11-13. Provider descriptions and published terms are attributed below. Published facts are attributed to their sources. Worked files are labeled as illustrative or as an inspected editorial source; neither implies commercial dataset clearance.

  1. AMI Corpus annotation layers and reliabilityUniversity of Edinburgh AMI · Reviewed 2026-11-13

    Primary research/publisher documentation reviewed for the adjacent factual reference. Benchmark availability and terms are distinct from a Fiund project; no hands-on corpus testing is reported.

  2. AMI Corpus annotation guidelinesUniversity of Edinburgh AMI · Reviewed 2026-11-13

    Primary research/publisher documentation reviewed for the adjacent factual reference. Benchmark availability and terms are distinct from a Fiund project; no hands-on corpus testing is reported.

  3. PodcastFillers dataset annotation vocabularyPodcastFillers authors · Reviewed 2026-11-13

    Primary research/publisher documentation reviewed for the adjacent factual reference. Benchmark availability and terms are distinct from a Fiund project; no hands-on corpus testing is reported.

  4. PodcastFillers audio and metadata license notesPodcastFillers authors · Reviewed 2026-11-13

    Primary research/publisher documentation reviewed for the adjacent factual reference. Benchmark availability and terms are distinct from a Fiund project; no hands-on corpus testing is reported.

  5. SPoRC: Mapping the Podcast EcosystemACL Anthology · Reviewed 2026-11-13

    Primary research/publisher documentation reviewed for the adjacent factual reference. Benchmark availability and terms are distinct from a Fiund project; no hands-on corpus testing is reported.

Sourcing with Fiund

Turn the requirements into a sourcing brief.

Send the model task, required source characteristics, target volume, annotations, intended use and review criteria. Fiund can assess sourcing options and the preparation needed. Availability, source documentation, permitted uses and pricing are confirmed for the proposed project.

Discuss podcast dialogue-act datasets: conversational function beyond speaker labels ↗

Build with Podcast Datasets

Start with the data
your model needs.

Share your use case. Get a focused next step.

Request a sample

Original studio and recording photography. See the media notes for source context.