Lithi Translation

Start with aligned source and target segments

For example, translate SEG-01–SEG-03 from en-US to fr-FR with glossary version A. Preserve each segment ID and {customer_id} placeholder. Keep SEG-03 unresolved when its referent is missing, rather than choosing a meaning from confidence alone.

Your agent defines the language, terminology and meaning a translation must preserve. It prepares permitted source units, protected text and any timing rules. Lithi returns aligned translations or explicit exceptions. Your agent checks identity, terminology, timing and unresolved context before accepting the result.

Your agent keeps the language decision and final review. The Kit prepares the supported translation request; participating Macs process the bounded segments, and your agent checks alignment, placeholders, terminology and timing.

Text, transcripts, and caption segments are supported source content within the Text & Data family; audio and video timing may remain part of the source evidence. The translation operation supplies a target language and output contract.

The operation is the exact supported capability. Agent guidance explains the process; it does not switch languages or grant access. The Translation pack covers translate-text, translate-aligned-segments, translate-transcript, translate-subtitles, and localize-corpus.

Before requesting a quote, resolve the supported language pair, operation, model revision, source format, glossary, terminology profile, and output contract. Do not assume every pair or asset is available. Keep source language, target language, recognition locale, subtitle locale, and optional translated audio separate. Translated text does not imply translated speech; subtitle translation does not burn captions into video or preserve rich layout unless the selected operation says it does.

Conceptual task brief: Translate source segments SEG-01–SEG-03 from en-US to fr-FR with glossary version A. Preserve segment IDs, {customer_id} placeholders, and source timing. Keep SEG-03 unresolved because its referent is missing. Return one aligned target or explicit exception per source segment. Accept only when protected text and identity counts match and every unsupported or missing segment remains visible.

The selected translation operation and supported integration map that brief to the permitted request and returned evidence. It does not silently switch languages or grant archive access. A human edit remains a new revision of the accepted result.

Preserve aligned identities and protected text

1. Freeze the source version, source segment or item IDs, glossary and terminology version, and output format.
2. Mark placeholders, markup, product names, and protected tokens before translation. Check count, identity, escaping, and placement afterward.
3. Preserve aligned target IDs and any confidence or human-review flags supplied by the operation. Keep missing and unsupported segments visible.
4. For transcripts, retain timing at source granularity. Sentence timing is not word alignment. For subtitles, reflow target text to the caption profile without changing source speech timing.
5. Review terminology, tone, context, and exceptions. A human edit creates a new revision; it does not rewrite the engine's original result.

These are illustrative review fixtures, not a universal schema or measured translation. Count source units by source identity and declared target language. Two target languages do not turn one source segment into two source inputs.

Handle timing, layout, and unsupported pairs

If a source segment is missing, unsupported, or fails, retain that state. If subtitle text overflows, revise the caption profile or send it to a separate layout operation. If a requested pair or rich document layout is unsupported, choose a supported alternative or stop for review; never silently switch language or claim equivalent output.

A changed glossary, source, language pair, caption profile, or timing requirement creates a new request version with its own quote and approval. A failed artifact download calls for recovery of the accepted artifact, not another translation submission.

Estimate translation work from the source

Tell your agent the result you need, the source type, its length or duration, the number of source items, and the target languages. The agent counts prepared text or extracted transcript content, chooses the supported translation operation, and carries one stage's output into the next. It uses a published assumption when no source measurement exists, so the result is an allowance until the translation is produced. Private source files are measured locally by the agent authorized to read them; a chat accepts only bounded counts or descriptors and cannot inspect the private file. For a prepared prompt, count the complete input once with the released tokenizer and chat template. For a PDF, count extracted text and page visuals separately, with page visuals included only when the selected operation reads them.

For an illustration based on two hours of audio followed by five text translations, assume 25 audio input tokens per second, 200 transcript output tokens per minute, a translation instruction allowance of 5% of each source, and equal-length translated output. The transcript is 24,000 output tokens (120 minutes × 200). The five translations use 126,000 input tokens (5 × 24,000, plus the 5% instruction allowance) and produce 120,000 output tokens. Including the audio stage, the workflow allowance is 306,000 input tokens and 144,000 output tokens, or 306,000 × USD 0.05 / 1,000,000 + 144,000 × USD 0.10 / 1,000,000 = USD 0.0297 — about 3 cents.

Reuse the accepted transcript output as the source for each target language. Each translation adds its own instruction allowance; do not treat the five languages as five new transcript generations or retokenize the transcript five times while dropping that overhead.

This is planning arithmetic under the stated assumptions, not measured output or an executable quote. The 5% instruction allowance applies to this illustration; it is not universal. Dubbing, burned captions, and video re-encoding remain separate operations. Use Get a free estimate after you have the source details, and treat the signed quote as the authority before paid work.

Worked example

Translate aligned segments with a pinned glossary

Freeze these en-US source segments, target fr-FR, and glossary version A:

Preserve segment identity so every accepted translation joins back to its source. Keep Compute unchanged as a protected product name and retain any placeholder with its original count and escaping. SEG-03 could refer to a battery or a billing action; confidence cannot supply that missing context.

Compare published prices

You can see what it costs to translate documents or a transcript with an AI model and with a translation service, using each provider's published list price. These tables compare published prices for the same work, not output quality. AI models bill in tokens and translation services bill in characters, so each row uses its own unit.

The first example translates 5,000 documents into 10 languages:

• 1,500 words per document, about 2,000 tokens each.
• 8,250 characters per document, 412.5 million characters in all, because translation services bill per character.
• 10 target languages, so 50,000 translations.

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. Each service bills in its own unit. These tables compare published prices for the same work, not output quality.

The second example translates the transcript of 500 hours of video into 10 languages. Its published example metering profile uses 25 audio input tokens per second and 200 transcript output tokens per minute, then adds a 5% instruction allowance to each source and assumes equal-length translated output. Those assumptions apply to this example; 5% is not a universal translation rule:

• 9,000 words per hour of speech, about 4.5 million words in all.
• 24.75 million characters per language, counting spaces.
• 10 target languages.

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. Each service bills in its own unit. These tables compare published prices for the same work, not output quality.

Use Audio & Video when source timing or media artifacts are part of the request and Documents & Search when citations or layout regions must survive translation. Continue with exception review before importing accepted language output into a customer workflow.