Transcript Format Ingestion

annotation_task_name: "Transcript Formats"
task_dir: "."
port: 8000

# Six instances, six different transcript formats. All of them render as the
# same speaker bubbles, because Potato normalizes every supported format to one
# turn model before display.
#
# Here the transcripts are inlined into the data file. They can equally be paths
# to sidecar files next to the media, which is usually what your ASR tool
# already produced:
#
#     "conversation": {"audio": "media/talk_01.mp3",
#                      "transcript": "media/talk_01.srt"}
#
# To build a data file from a folder of ASR output instead:
#
#     potato transcripts ./whisper_out --media-dir ./audio -o data/items.json
#     potato transcripts ./whisper_out --dry-run     # what did it detect?

data_files:
  - data/sample-data.json

item_properties:
  id_key: id
  text_key: title

instance_display:
  layout:
    direction: vertical
    gap: 16px
  fields:
    - key: title
      type: text
      label: "Format"

    - key: source_format
      type: text
      label: "Where this came from"

    - key: conversation
      type: audio_dialogue
      label: "Transcript"
      span_target: true
      display_options:
        audio_key: audio
        scroll_height: 420px
        show_timestamps: true
        # "auto" reads the field as a sidecar path when it looks like one.
        # These transcripts are inlined, so nothing is read from disk.
        transcript_is_path: auto
        # No fixed roster: these files use different speaker conventions
        # (Host/Guest, Interviewer/Participant, speaker_0/speaker_1, tier
        # names), so speakers get deterministic colors and alternating sides.
        allow_speaker_assignment: auto

annotation_schemes:
  # Whether the transcript survived ingestion intact, which is the judgment you
  # actually want a human to make before trusting a corpus.
  - annotation_type: radio
    name: ingestion_quality
    description: "Did this transcript come through correctly?"
    labels:
      - name: "clean"
        tooltip: "Timings, speakers, and text all look right"
      - name: "timing_off"
        tooltip: "Turns are misaligned with the audio"
      - name: "speakers_wrong"
        tooltip: "Turns are attributed to the wrong speaker, or not attributed"
      - name: "text_garbled"
        tooltip: "Text is missing, truncated, or mangled"

  # Per-turn labeling, bound to every turn of the conversation field.
  - annotation_type: radio
    name: turn_type
    description: "What is this turn doing?"
    labels:
      - question
      - answer
      - aside
    turn_level: true
    turn_binding:
      field: conversation

  # Spans work over normalized turns regardless of the source format.
  - annotation_type: span
    name: highlights
    description: "Highlight anything an annotator would need to check against the audio"
    target_field: conversation
    labels:
      - name: unclear
        key_value: "u"
      - name: likely_asr_error
        key_value: "e"

output_annotation_dir: "annotation_output/"
export_annotation_format: "jsonl"