Skip to content

desktop-app.md

Latest commit

 

History

History
62 lines (41 loc) · 3.61 KB

desktop-app.md

File metadata and controls

62 lines (41 loc) · 3.61 KB
Raw

Vernacula Desktop

Vernacula-Desktop is the Avalonia-based GUI for running the Vernacula pipeline on your own machine. It wraps the full stack — diarization, ASR, optional language identification, and optional KenLM fusion — behind a job queue and transcript editor.

Vernacula-Desktop

Vernacula-Desktop converts audio files into accurate, multi-speaker transcripts — entirely on your own computer. No cloud uploads, no subscriptions, no privacy concerns. Works on Linux, Mac, and Windows (Android, iOS, and WebAssembly are untested).

Powered by NVIDIA's Parakeet TDT v3 and Sortformer by default, with optional pluggable backends (Cohere Transcribe, Qwen3-ASR, VibeVoice-ASR). Parakeet v3 posts a Word Error Rate of 4.85 on Google's FLEURS benchmark — among the best available anywhere. Most modern computers will transcribe one hour of audio in about five minutes. GPU-accelerated systems are significantly faster.

Demo

vernacula-desktop_usage_demo.mp4

Screenshots

Jobs view — queue and manage transcription jobs

Jobs view

Results view — review the completed transcript with speaker labels and timestamps

Results view

Transcript editor — correct text, adjust timing, and verify segments with audio playback

Transcript editor view

Highlights

  • Local, private transcription — audio never leaves your computer
  • Multi-speaker detection — identifies and labels up to four concurrent speakers
  • No audio length limits — streaming and segmentation handle indefinite file lengths
  • Job queue — pause and resume long transcription jobs
  • Automatic punctuation and capitalization from the acoustic model
  • Transcript editor with confidence colouring, audio playback, word-level highlighting, and segment editing
  • Word-level timestamps (real TDT duration-head values for Parakeet, synthesized for others)
  • Beam search + shallow KenLM fusion for Parakeet — opt-in domain biasing via the Settings → Language model dropdown
  • Language identification (VoxLingua107, optional) with file-level or per-segment modes
  • Wide format support — common audio formats plus MP4, MOV, MKV, AVI, WMV, FLV, MTS, and more
  • Export to XLSX, CSV, JSON, SRT, Markdown, DOCX, and SQLite
  • Full analysis data in SQLite with per-token durations, logprobs, and speaker labels
  • GPU acceleration via CUDA (DirectML on Windows), with automatic CPU fallback
  • Parakeet v3 covers 25 languages: English, French, German, Spanish, Portuguese, Italian, Dutch, Polish, Russian, Ukrainian, Czech, Slovak, Romanian, Hungarian, Bulgarian, Croatian, Slovenian, Greek, Swedish, Danish, Finnish, Estonian, Latvian, Lithuanian, and Maltese
  • Qwen3-ASR and Cohere Transcribe backends add ~30 and ~15 additional language options respectively

Built with Avalonia UI — runs on any desktop environment.

Running from source

cd src/Vernacula.Avalonia
dotnet run

For packaged installs on Linux, see Installation. For custom build configurations, see Building from source.

Models

The Avalonia app surfaces each model repo through Settings → Models and handles downloads, resuming, and hash verification. See Models for the full catalogue.