Skip to main content
Every transcribed episode has three transcript views, each tuned for a different use case:
  • Dialogue — speaker-attributed lines with sentence-level timestamps. The default. Best for chat-style UIs, LLM context windows, and most reading.
  • Words — every word individually timestamped. Best for karaoke-style highlighting, precision audio editing, and word-aligned search.
  • Mentions — lines surrounding mentions of a specific entity, with is_mention flags. Best for “what did they say about X?” workflows.
Dialogue transcripts are also reachable scoped to a single segment or clip.
Available to MCP agents via particle_podcast_get_episode with include: ["transcript"] (and optional transcript_format, transcript_speaker, transcript_start, transcript_end). For mention-style windows, use particle_podcast_find_mentions.

​
Dialogue transcript

Response (truncated)
Lines are ordered, 1-indexed, and include speaker name and role. Note that the first lines of many episodes are sponsorship reads — fetch /segments to see which time ranges are tagged AD if you want to skip them. See segments.

​
Output formats

Use the format query parameter:
Structured JSON with speaker attribution, roles, and timestamps per line.
Best for: building conversation UIs, speaker analysis, programmatic processing.

​
Filter by speaker or time range

Extract everything one person said:
Extract a time range:
Combine — exactly what one person said during one segment:

​
Word-level transcript

For per-word timing — karaoke-style highlighting, precision editing, word-aligned search:
Response (truncated)
Word-level transcripts for long episodes can be very large — a feature-length episode runs to tens of thousands of entries. Pass start and end to clip a time range, limit to bound the page size, and exclude_spacing=true if you don’t need the inter-word whitespace tokens.

​
Query parameters

​
Spacing tokens

Roughly half of the entries in a typical word transcript have type:"spacing" and text:" " — these represent the silence between spoken words, with their own start_seconds/end_seconds. They’re useful when timing matters: caption/subtitle UIs that need precise pause durations, karaoke-style word highlighting against playback, or reconstructing inter-word silences for audio alignment. NLP, search, and LLM consumers should pass exclude_spacing=true to skip them — it roughly halves the payload.

​
Pagination

When you pass limit, the response includes a cursor and has_more: true until the last page:
The speaker field is the identified speaker name (e.g. Scott Galloway). When a speaker hasn’t been resolved to a name, the raw STT label (speaker_0, speaker_1, …) is returned instead.

​
Transcript mentions

The most useful transcript view when you care about a specific person, company, or topic. Returns the dialogue lines around every mention of an entity in one episode, with is_mention: true on the lines that actually contain the mention and surrounding lines for context.
Response (truncated)
Use start_seconds / end_seconds to deep-link into the audio, or feed the line text into an LLM with full speaker context.

​
Chunking semantics

The endpoint groups transcript lines into context windows:
  • context_lines (default 2, max 20) is the radius of each window. A mention on line i produces the closed range [i - context_lines, i + context_lines], clamped at the transcript boundaries — so a single mention yields up to 2 * context_lines + 1 lines.
  • Adjacent or overlapping windows merge into a single window. Two mentions within context_lines of each other produce one entry with multiple is_mention: true lines, not two entries.
  • total_mention_count is the unfiltered count of dialogue lines containing at least one mention of the entity in the episode. It is independent of pagination, and equals the sum of is_mention=true flags across the un-paginated mentions[].
  • Mention matching is a case-sensitive substring match against mention_variants (the entity’s canonical name plus any annotated aliases).

​
Pagination

When entity_id is set, exactly one entity is returned and its mentions[] is paged. When entity_id is omitted, each entity in the page returns all of its mentions; deep-paging through a single entity’s mentions requires re-issuing with that entity_id. limit defaults to 25 and is capped at 100. Pass the cursor from a previous response to fetch the next page; has_more indicates whether more results exist.

​
Segment and clip transcripts

Segments and clips have their own transcript endpoints scoped to that time range:
Both accept the format query parameter (dialogue, text, or srt).