Lithi Audio & Video

Start with the clip decision

For episode EP-01, ask for a candidate covering 4–12 seconds and keep the qualification spoken at 8–12 seconds. Return the source time, transcript or frame evidence and review state. A 4–8 second cut that omits the qualification needs rejection or extension; an agent cannot make it publishable by shortening the interval.

Your agent defines the editorial question and selects permitted source material. It prepares the source version, intervals and review rule for the chosen media operation. Lithi returns a source-linked transcript, scene map, frame, caption or candidate. Your agent checks timing, coverage and qualifications before handing the material to an editor.

Your agent keeps the editorial question and publishing decision. The Kit prepares the supported media request; participating Macs process the bounded interval, and your agent reviews timing, coverage and qualifications.

Recordings, meetings, podcasts, clips, and video libraries belong to the Audio & Video information family. Start with one immutable source or permitted proxy and the decision it needs to support. The operation determines whether the result is a transcript, scene map, frame, caption, or candidate.

The operation is the exact supported capability. This guide describes how to prepare and review its source-linked output. Agent guidance does not fetch a source or publish media.

The Creator & Media pack covers transcribe-media, make-subtitles, analyze-video, extract-keyframes, detect-scenes, clip-candidates, thumbnail-candidates, make-vertical-short, and prepare-content-pack. An overview can provide a scene map and representative previews; add detail only to selected intervals. State features, sampling, locale, model or profile, maximum gaps, resolution, and artifact limits before quoting.

Conceptual task brief: For source EP-01, review the 4–12 second interval for a clip candidate. The permitted evidence is the source version, transcript ranges, and decoded frame or scene evidence for that interval. Keep the qualification at 8–12 seconds visible; abstain when timing or source identity is uncertain. Return the candidate with source time, supporting transcript or frame evidence, and review state. Accept only when the interval is mapped to the source and no material qualification is cut.

The selected media operation and supported integration map that brief to the permitted request and returned evidence. The editor remains responsible for the final selection and production decision.

Keep every result tied to the source

1. Bind each transcript segment, scene interval, frame, observation, caption cue, or candidate to the source asset and source version.
2. Preserve rational time values, timescales, stream identity, requested time, actual decoded time, and any mapping to a proxy or final edit.
3. Keep the original transcript, display cleanup, correction, translation, and creative copy separate. A transcript does not imply captions, and caption data does not imply a rendered video.
4. Specify recognition locale, target translation locale, subtitle profile, and each optional render or generation stage separately.
5. Inspect the result package, coverage, exceptions, and artifact digests before making an editorial selection.

For a format change, your agent can check current operation discovery for a supported video transcode path for EP-01 and the desired container and codecs. If available, it can prepare a separate request and review the returned video's decoding, duration and synchronization against the source. An unsupported path remains an exception; a transcode does not make editorial selections.

make-vertical-short refuses landscape or wide input. Use a portrait source; customer- or editor-controlled reframing is a separate decision. A thumbnail candidate is an image for review, not a prediction of clicks.

Review evidence before selecting material

These rows are synthetic review targets, not fixed API fields or measured media output. Check actual frame time against requested time. A contact sheet is a preview map; selecting a tile does not select its crop as the final asset.

Reject a candidate that cuts away a material qualification, has insufficient interval coverage, or uses an unsupported coordinate or time mapping. Keep NOT_RUN, UNSUPPORTED, FAILED, and INSUFFICIENT_EVIDENCE distinct from observed zeroes. Anonymous face or speaker track IDs do not identify real people.

Recover a source or artifact failure

If a source cannot be decoded, keep it as an exception instead of reporting no media. If a download fails after an accepted result, retrieve that artifact again. If the source edit, proxy, caption style, language pair, or requested render changes, preserve the earlier result and prepare a new quoted request. Do not rerun inference merely because delivery failed.

Estimate the media workflow

Before requesting a quote, give your agent the result you want, the source type, duration, frame dimensions, sampling or pass count, and target languages. The agent counts decoded audio duration and admitted image or video dimensions, chooses the supported operations, and links each stage's output to the next. An estimate uses a published assumption when the source has not been measured; it remains an allowance until the work is produced. Private media is measured locally by the agent authorized to read it; a chat accepts only the bounded counts or descriptors the agent supplies and does not inspect the private file. Any prepared prompt is counted once with the released tokenizer and chat template, without counting an embedded file a second time.

Here is an illustration for two hours of audio, its transcript, and five text translations. Under these assumptions, audio uses 25 input tokens per second, transcript output uses 200 tokens per minute, each translation adds an instruction allowance of 5% of its source, and each translated output is the same length as its source:

The accepted 24,000-token transcript is the source reused for each target language. Each translation has its own input allowance and instruction overhead; it is not five new transcript generations, and the source is not counted five times without that overhead.

At the customer tariff, that illustration is 306,000 × USD 0.05 / 1,000,000 + 144,000 × USD 0.10 / 1,000,000 = USD 0.0297, or about 3 cents. It is planning arithmetic under the stated assumptions, not measured output or an executable quote. The 5% instruction allowance is specific to this illustration; it is not a universal translation rule. Dubbing, burned captions, and video re-encoding are separate operations and are not included in this total. Use Get a free estimate to describe the workload, then review the signed quote before approving paid work.

Worked example

Preserve a qualification in a clip candidate

For episode EP-01, source version A, the synthetic timeline is:

0–4 seconds: Today we discuss reusable packaging.
4–8 seconds: This box can be reused.
8–12 seconds: But not after it becomes wet or damaged.

The transcript has three source-linked ranges. A clip candidate covering 4–12 seconds keeps the qualification; a 4–8 second candidate drops it and must be rejected or extended. Draft copy such as “Reuse packaging only while it remains suitable” is still editorial material. A usable transcript does not make every clip candidate publishable.

Compare published prices

You can see what one job costs at published list prices, step by step. These tables compare published prices for the same work, not output quality.

The example is 500 hours of video: 30,000 minutes, or 1,800,000 seconds of speech and picture. The published example metering profile uses 25 audio input tokens per second and 200 transcript output tokens per minute. Its ten-language translation example adds a 5% instruction allowance to each source and assumes equal-length translated output. Those are example assumptions, not a universal translation rule; a signed quote sets the real charge. Some services have no managed equivalent for a step, and the table says which.

Transcribe the speech

Published list prices checked 5 October 2026.

Each provider counts tokens with its own tokenizer, so the same text can be a different number of tokens. Rows with a note count input with that provider's own published rule. Each service bills in its own unit. Lithi figures use the published example metering profile; a signed quote sets the real charge. These tables compare published prices for the same work, not output quality.

Find scenes and keyframes

Published list prices checked 5 October 2026.

Each provider counts tokens with its own tokenizer, so the same text can be a different number of tokens. Rows with a note count input with that provider's own published rule. Each service bills in its own unit. Lithi figures use the published example metering profile; a signed quote sets the real charge. These tables compare published prices for the same work, not output quality.

Choose clips from the transcript and scene list

Published list prices checked 5 October 2026.

Each provider counts tokens with its own tokenizer, so the same text can be a different number of tokens. Rows without a note apply each published rate to the same token counts. These tables compare published prices for the same work, not output quality.

Whole job

The whole job is transcription, translation into 10 languages, scene detection and clip selection.

Published list prices checked 5 October 2026.

AWS and Google Cloud sell no clip-selection service, so their totals cover transcription, translation and scene finding only. gpt-6-luna and Claude Haiku 4.5 publish no audio or video input rate, so they have no total for all four steps. Each service bills in its own unit. Lithi figures use the published example metering profile; a signed quote sets the real charge. These tables compare published prices for the same work, not output quality.

Use Translation when a language pair is part of the request and Image Analysis for image evidence. Continue with artifact verification before handing selected material to an editor.