Preview Note

Create notes faster, at a lower cost

For one hour of footage, manual preview-note writing takes at least 4 hours; Preview Note takes 15 minutes. AI automatically organizes speakers, dialogue text, and scene descriptions. Your team can start by reviewing the completed draft and move into editing sooner.

Note generation costs USD 0.15 per minute of footage. Review speakers, dialogue text, and scene descriptions on one timeline, make corrections, and use the notes as your team’s editing reference.

Preview notes in your broadcast format

The input is one raw video file (optionally the production brief for that episode). The output is a DOCX in the broadcaster’s own format, plus the same content as Markdown.

The following example illustrates the document structure.

Preview note: 0628-1-2 — <item title>
================================================================
Cast : Talent A, Talent B, Crew (PD / writer)
================================================================
T/C     Content

[0004] Morning at the ranch, feed mixing instructions
0004    Talent A   Ah, so there's one leader. Steadier today.
0006               Two people stand inside the pen holding feed.
0009    Crew       There is a leader, but apparently not a clever one.

Line by line, the structure is visible:

• Timecode — milliseconds from the start, converted to broadcast format
• Speakers — distinguish who is speaking, with names where confirmed
• Dialogue text — transcribed speech aligned with speakers and timecodes
• Scene descriptions — written descriptions of people’s actions and scene context
• Section heading — a one-line summary at each scene or topic change
• Header — program, cast and background pulled from the brief

Speakers, dialogue text, and scene descriptions share one timeline. Use timecodes and headings to find the right footage, then edit and share the notes in your team’s document format.

Combining dialogue, scenes, and production context

Video ─┬─▶ ① global word-level STT + diarization ──▶ dialogue + voice embeddings
       └─▶ ② segmented scene detection + keyframes + OCR ──▶ description + on-screen text
                    ③ brief parsing ──▶ proper nouns, cast context
                    ④ spelling correction + voiceprint name matching
                    ⑤ timeline merge + section headings + DOCX render

Why transcription has to be global

The Speaker 1 / Speaker 2 labels diarization produces are local identifiers, valid only within one processing unit. Processing a long recording in separate chunks can assign different labels to the same person, making the notes harder to review.

So the pipeline splits: audio is processed whole in one pass (speaker consistency), while scene detection, description and OCR run per segment (a scene is a local unit anyway).

Speakers assigned to words, not segments

The usual STT + diarization combination assigns a whole 5–20 second segment to whichever speaker overlaps it most. On a real shoot that collapses:

“…so this is how we mix the feed” (talent) → “Hold on, long shot here” (PD) → “Ah, right” (talent)

Put those three utterances in one segment and all three become the talent’s line — a directing instruction turns into on-screen dialogue. In a preview note that is fatal.

Preview Note assigns a speaker to each word and breaks lines at speaker changes, with a 4-second silence gap and a 7-second maximum line length in this processing configuration. Preview-note layout and final subtitle delivery specifications are confirmed separately for each project.

The production brief as context

This is where the biggest gap with general-purpose tools appears. Broadcast production has a brief: program name, cast names and roles, locations, product names. A human previewer works with it open beside them. Machines did not.

The brief is parsed from PDF or DOCX and structured, and the extracted proper-noun list does two jobs.

1. It grounds spelling correction. Speech recognition mis-spaces and mis-spells business names; the glossary fixes them. Correction runs under strict constraints — do not change meaning, do not add or remove content, return exactly the same number of items as the input. If the count doesn’t match, that batch keeps the original text.
2. It identifies which item the footage belongs to. Filename numbers are often camera or tape numbers rather than item numbers, so matching is content-based: count how often each item’s proper nouns appear in the transcript. Simple counting — and far more reliable than filename convention.

Carrying speaker labels beyond one video

Speaker labels are assigned within each recording. For a series, 256-dimension voice embeddings accumulate into per-person profiles and are compared by cosine similarity. Names reviewed by a person in one episode support matching in later episodes. Uncertain matches retain an anonymous speaker label for review.

The LLM proposes, deterministic code constrains

A flat timeline thousands of lines long cannot be read, so section headings are attached at transition points. But timestamps the model returns are snapped to the start time of a line that actually exists, and a minimum interval between headings is enforced in code.

LLMs do not reliably obey instructions like “not too frequently.” So we ask in the prompt and enforce in code — a principle that runs through the whole system.

What we confirm with you up front

Vision captioning has no domain knowledge. It describes from a single keyframe, so unusual subject matter can be mislabeled. Injecting brief information into the captioning prompt is the next improvement.

Named speakers are an aid, not a guarantee. Matching is deliberately conservative and falls back to Speaker N when ambiguous. The tool assumes previewer review.

On-screen text (OCR) is opt-in. Burned-in timecode and camera UI mix in easily, so it is off by default and enabled for projects that need meaningful on-screen text such as captions and signage.

Review the notes and move into editing

Dialogue text, scene descriptions, and speaker information are saved by stage to support resuming interrupted work. Receive the result in your team’s broadcast DOCX format and Markdown for editing and sharing.

The deliverable is a draft for the production team to compare with the source. Review overlapping dialogue, names, and specialist scene descriptions before completing the broadcast document. Adoption uses actual footage and the team’s document template to assess quality and working time.