Paste a link, get the text back

Scope

How it’s built

1. 1Desktop shell

   • Tauri 2 (Rust) + React
   • First run fetches only what it needs, about 222 MB
   • Runtime packs verified by SHA-256

2. 2Local service

   • FastAPI, launched with the app
   • Listens only on a random local port
   • A short-lived token per launch
   • Progress streamed over SSE

3. 3Get the text

   • Platform subtitles → page text → transcription
   • If subtitles or text exist, skip transcription
   • Video: fetch the audio first and deliver the text; the original video follows in the background

4. 4Transcribe

   • SenseVoice on CPU by default
   • CUDA on an NVIDIA GPU, when there is one
   • CPU and GPU runtimes in separate processes

There is only the Windows app: fetching and transcribing both happen on your own computer.

Trade-offs

• SenseVoice by default

  WhyOn a CPU, SenseVoice runs at a real-time factor of about 0.035 (an hour of audio in about two minutes) versus about 2.01 for Faster-Whisper medium. On the same 72 hand-transcribed clips, both on a GPU, three rounds: Mandarin character error rate 12.75% for SenseVoice and 37.50% for Whisper medium, in 23 s against 138 s

  Given upA bigger default model

• Use subtitles and page text when they exist

  WhyThe fastest transcription is the one you skip

  Given upTranscribe everything

• Text first, original video in the background

  WhyAll 4 old timeouts were waiting for the whole original video (a 260 MB sample) before transcription; with the change, the same links ran three rounds with no timeouts

  Given upWait for the video, then transcribe

• The background video keeps the original quality

  WhyThe split streams alone are faster, but the same video came out at 854×462 instead of the original 1280×692

  Given upTrading resolution for speed

• Read Xiaohongshu’s structured page data directly

  WhyNo full-page HTML: transfer fell from 521 KB to 2 KB and total time by 30–40%

  Given upFetch the page, then parse it

• Do three things only

  WhyAccounts, payments and workspaces turn a tool into a platform

  Given upBecoming a platform

Timeline

1. Step 1Get it running in the command line: install FFmpeg, download a transcription model, until the first text came out
2. Step 2Give it a good interface, modelled on Apple’s design
3. Step 3Pull images and text from Xiaohongshu and other platforms too, without their watermarks
4. SinceRework it again and again until it was stable
5. 2026.10.23v2.3.0: AI proofreading, GPU acceleration, an estimate of the time left; video text first, the original in the background

Evidence

• A live smoke test with real samples on 2026-08-11: Bilibili, Xiaoyuzhou, Douyin and WeChat articles passed; a Xiaohongshu post with 15 images and 14 Live clips passed.
• 2026-10-21: five platforms, 60 links, three rounds each. 137 of 180 runs delivered in full and 4 timed out, so 2.3 was held back.
• 2026-10-22, after the fixes, the same links over three rounds: 141 delivered in full, no timeouts. Bilibili 57/57, WeChat 30/30, Xiaoyuzhou 27/30, Douyin 24/30, Xiaohongshu 3/33; every failure stays in the count.
• Accuracy: on the public ASCEND corpus (hand-transcribed Hong Kong conversations in Chinese and English), the GPU build of SenseVoice scored a Mandarin character error rate of 9.31%, an English word error rate of 23.34% and 15.52% on mixed speech (5,735 clips, three rounds each).
• The v2.3.0 installer was accepted on his own PC and then on a clean Windows machine; the file released is that same file, and a download from the website matches its SHA-256.
• Each platform’s status is published with the date of its last check, not as a permanent promise.
• Installers, runtime packs and release files all carry SHA-256; GitHub Actions also produces an SBOM and build provenance.