Written text in. Sign language out.

Poze translates written sentences into animated signing. It plans the sentence the way a signer would, picks recorded signs from real dictionaries, and moves a 3D signer with natural timing, facial movement and use of space. We start with Turkish Sign Language (TİD).

    • Dictionary sign
    • Chosen by context
    • Substitute
    • Fingerspelled
    • Space and direction
    Loading the signer

    Choose a sentence. Each one shows a different part of the system.

    Three stages, each one you can inspect.

    A sign language is not a word-for-word copy of the spoken language. Word order, space and movement carry grammar. Poze keeps each decision visible, so an error can be traced to the stage that made it. The panels below follow the sentence playing above.

    1

    Interpret

    Claude Sonnet 5 reads the sentence and writes a sign plan: TİD word order, negation, questions, names, and which referents get a place in signing space.

    Input

    Sign plan

    2

    Select signs

    Each planned sign is matched to recorded dictionary signs. Context picks the right sense. When no sign fits, Claude can reject every candidate and the word is fingerspelled.

    3

    Animate

    Body, hand and face motion tracked from the source videos is joined with retimed transitions, kept inside human joint ranges, and bent toward points in space for directional verbs.

    Try your own sentence.

    Write a Turkish sentence and the full pipeline runs live: Claude Sonnet 5 writes the sign plan, the server selects signs and builds the motion, and your browser animates the signer.

    Checking the live server

    • Sign plan
    • Select signs
    • Face
    • Motion
    Your signing appears hereWrite a sentence and press Translate to TİD. A short sentence takes about 20 seconds. If the server has been idle, the first one takes about a minute longer while it starts.

      Built once. Extended one sign language at a time.

      There are more than 300 sign languages. Most have no automatic translation at all. Poze separates what every sign language shares from what each one needs, so a new language means new recordings and new rules, not a new system.

      Shared across languages

      • Sign plan format. A gloss sequence with markers for negation, questions, reduplication and spatial reference.
      • Retrieval and fallback. Sense selection from sentence context, synonym and root-form matching, and a fingerspelling fallback when no sign fits.
      • Motion engine. Pose tracks from video, retimed transitions, joint limits, hand contact, face and mouth movement, and spatial agreement.
      • Evaluation. Automatic gloss and pose metrics, plus study tools for Deaf signers and interpreters.

      Added for each language

      • Recordings. A dictionary of signs from native signers, with consent and licences for reuse.
      • Linguistic rules. Word order, negation and question marking, verb agreement and the manual alphabet.
      • Examples. Parallel sentences that teach the interpretation step the target grammar.
      • Validation. Ratings from signers of that language before anything is used in public.

      Measured, not assumed.

      Poze grew out of academic research on sign language and language models. We test every stage and we test with the people the system is for.

      371held-out Turkish sentences in the automatic test set
      99.9%of planned signs realised in the video on that set (5,056 of 5,062)
      11model setups compared for the sign plan, open and closed
      3conditions in our signer study: Poze, Poze with expert plans, and a real signer

      Built on Claude.

      Claude Sonnet 5 handles the two steps that need understanding of meaning. Every other step is deterministic and traceable, so a language model decision is always a named, reviewable step.

      Prompts carry TİD-specific instructions and worked examples, and results are cached by sentence and prompt version for fast, low-cost repeat requests.

      1. Writing the sign planTurkish sentence to an ordered TİD gloss sequence with negation, question and spatial markers.
      2. Judging substitutesWhen a sign is missing, Claude compares candidate signs and their definitions with the sentence, then accepts one or rejects them all in favour of fingerspelling.
      3. Next: annotation supportConsistency checks on our own sign recordings, with every linguistic decision reviewed by a person.