Qwen Transcribe

Local speech-to-text with Qwen3-ASR-1.7B: live dictation from any microphone, plus audio/video file transcription, with a searchable history. Everything runs on this machine; nothing is sent anywhere.

You need Windows 10/11, an NVIDIA GPU with ~6 GB of free VRAM and a recent driver (tested on an RTX 5090), Miniconda, and ~12 GB of disk. Run these in an Anaconda Prompt:

git clone https://github.com/mkecman/qwen-transcribe.git
cd qwen-transcribe

:: 1. Python 3.12 + ffmpeg, in a conda env inside the project folder
conda create -p .\env -c conda-forge --override-channels python=3.12 ffmpeg

:: 2. Python packages (PyTorch with CUDA 12.8, Qwen3-ASR, Silero VAD, FastAPI)
env\python.exe -m pip install -r requirements.txt

:: 3. The model (~4.4 GB) into models\
env\Scripts\hf.exe download Qwen/Qwen3-ASR-1.7B --local-dir models\Qwen3-ASR-1.7B

Then run start.bat.

Double-click start.bat. It loads the model (~10 s) and opens http://127.0.0.1:8765 in your browser. Close the console window to stop it.

• Dictate: pick a microphone in the bottom bar, then click Record (or press Ctrl+Space). Text appears as you speak; the grey tail is still being refined. Press Enter (or the ¶ button) to start a new paragraph. Stop with the same button or Ctrl+Space.
• Continue a session: open it from the history and press Continue. New text is appended as a new paragraph.
• Files: drag audio or video files anywhere onto the page, or use the upload button. Anything ffmpeg reads works (mp3, m4a, wav, ogg, flac, mp4, …). A 6-minute recording takes about 6 seconds.
• Edit and copy: transcripts are editable (autosaved). Copy puts the whole text on the clipboard. The ⋯ menu has Download .txt, Re-transcribe (from the saved audio), and Delete.
• Settings (sliders icon, bottom left): default language, a vocabulary list for names and jargon, filler-word cleanup, how long a pause finalizes text, auto-paragraphs, mic processing, whether to keep audio, and the theme.

• app/engine.py: loads the model once and runs all GPU work on one worker thread, with priorities (live text first, file jobs last). Decoding replays a captured CUDA graph per token instead of launching ~2,000 small kernels. On Windows this made generation ~4–8× faster, at ~5–6 ms per token, with output identical to the reference implementation.
• app/segmenter.py: Silero VAD splits audio at pauses. File uploads use a batched pass that is 6× faster.
• app/live.py: live dictation over a WebSocket. Utterances are grouped into ~40–55 s windows. At each pause the window is re-read, with its earlier transcript forced as the start of the answer, so a sentence spanning a thinking pause comes out whole and only the new part is decoded.
• app/files.py: file jobs. Speech is grouped into 40–60 s chunks, cut at the longest pause.
• Seams: when a piece of audio ends, the model ends its text as if the speech had stopped. It adds a period and closes any open quote, which leads to sentences like "which is essentially. Funneling into…". So each chunk (and each new live window) is decoded together with the previous one, forcing the previous text minus its final punctuation. The model then punctuates the seam knowing what comes next.
• app/textutil.py: filler and stutter cleanup. It also repairs quotation marks: speech has no audible closing quote, so a quote the model never closes, or one that runs implausibly long, loses its marks instead of swallowing the narration that follows.
• web/: the UI, plain HTML/CSS/JS with no build step. worklet.js captures 16 kHz PCM from the mic.

Deleting the project folder removes everything. Nothing is installed outside it.

• Serbian is not supported by Qwen3-ASR. The language list in the app is what the model supports (30 languages, plus auto-detect).
• The vocabulary setting is passed to the model as context. It fixes spellings of names and jargon (e.g. "tepesquite" became "Tepezcohuite" once that word was listed), but keep it to a short list of terms: long text there can leak into the transcript.
• The server listens on 127.0.0.1 only, and rejects writes coming from other websites.
• Environment variables: QT_PORT (default 8765), QT_DATA_DIR, QT_MODEL_PATH.
• Linux should work too (env/bin/python -m app; it falls back to the system ffmpeg), but it's untested.

• Qwen3-ASR by the Alibaba Qwen team (Apache-2.0): the model and the qwen-asr package. The prefix-rollback continuation follows its streaming implementation.
• Silero VAD (MIT): speech detection.
• Icons adapted from Lucide (ISC).