IA on-device · WebGPU

Gere legendas e traduza no dispositivo.

Transforme qualquer vídeo ou áudio em legendas SRT, VTT ou ASS sincronizadas — com editor em nível de palavra e um segundo idioma, tudo no navegador.

Detectando GPU…
Saída bilíngue
100% privado
Por que Estúdio de Legendas

Legendas de qualidade broadcast. Zero uploads.

Transcrição Whisper, tradução Marian e um editor de legendas de verdade, tudo local.

100% privado

Seu material nunca sai do dispositivo. Decodificação, transcrição e tradução acontecem localmente.

Acelerado com WebGPU

A transcrição roda na sua GPU quando disponível, com fallback em WebAssembly.

Tempo por palavra

Modelos Whisper com timestamp fornecem tempos por palavra para legendas justas e exportação karaokê.

Saída bilíngue

Traduza para um segundo idioma e exporte as duas faixas empilhadas na ordem que preferir.

Como funciona

Do material às legendas.

  1. 01

    Solte um vídeo ou áudio

    Arraste MP4, WebM, MP3, WAV ou M4A — decodificado localmente, nada é enviado.

  2. 02

    O Whisper escreve as legendas

    A fala é transcrita em janelas sobrepostas e depois normalizada pelas regras de transmissão.

  3. 03

    Edite, traduza, exporte

    Corrija as legendas no estúdio, adicione uma tradução e baixe SRT, VTT, ASS, TXT ou JSON.

Complete guide

About the subtitle generator

A free video-to-subtitle tool that listens to your audio with Whisper, writes timed cues, and can translate them into a second language — all inside the browser. It is built for creators, teachers, journalists and anyone captioning interviews, lectures, screen recordings or vlogs without handing raw footage to a cloud service.

How it works

The audio track is decoded with the Web Audio API and resampled to 16 kHz mono, then transcribed by a Whisper model running through Transformers.js. The file is walked in overlapping 30-second windows, so long recordings stream progress as they go and word-level timestamps stay accurate across window boundaries. Because the tool uses the timestamped Whisper builds, every cue also carries per-word timings, which drive the karaoke preview and the ASS karaoke export.

For translation, compact Marian (OPUS-MT) models are used, pivoting through English when the source and target have no direct model. Both the transcription and the translation models download once and are then cached, so a second run of the same language pair works offline.

Editing and quality

Generated cues are not set in stone. The studio pairs a live preview — with a broadcast-style subtitle overlay and a waveform timeline you can scrub — with a cue editor that supports inline text edits, splitting at the playhead, merging, duplicating, search & replace across both tracks, global time shifting, and full undo/redo. A quality panel flags every cue whose duration, reading speed, line length or overlap breaks normal subtitle conventions, and one click re-normalises the whole document.

Finished subtitles export as SRT, WebVTT, ASS, plain text or structured JSON, in the original language, the translation, or both stacked in either reading order.

Limits & requirements

The first run downloads model weights, which is the slowest step; afterwards they are cached. WebGPU (Chrome or Edge) makes transcription several times faster, and Safari or Firefox fall back to WebAssembly. Translation uses compact open models that are strong for general conversation and narration but should not be trusted for legal, medical or otherwise high-stakes wording. Very long recordings can exhaust memory on low-RAM devices; splitting the file is the workaround.

Privacy

Cloud captioning services receive your unedited footage and your scripts. Here nothing is uploaded: the video is decoded, transcribed and translated on your own machine, and the only files that leave the tab are the subtitles you choose to download.

Suporte

Perguntas, respondidas.

Não. O arquivo é decodificado, transcrito e traduzido no seu navegador. Só o arquivo de legendas que você baixar sai da aba.