Skip to content

Releases: BenPohlBasel/LocalTranscript

LocalTranscript 0.1.0

25 Apr 22:20
@BenPohlBasel BenPohlBasel
v0.1.0
7cc74aa

Choose a tag to compare

View all tags
LocalTranscript 0.1.0 Latest
Latest

First public release of LocalTranscript — fully offline audio transcription with speaker diarization for macOS Apple Silicon.

Install

  1. Download LocalTranscript-0.1.0-arm64.dmg below
  2. Mount, drag LocalTranscript.app into /Applications
  3. Right-click → Open → Bestätigen on first launch (Gatekeeper, the build is not Apple-notarized)
  4. The app asks once where transcripts should be stored. Default: ~/Documents/LocalTranscript. Each batch creates a yymmdd_hhmm/ subfolder containing the original audio plus VTT, TXT and CSV.

What's bundled

  • Transcription: whisper.cpp with Apple Metal GPU, OpenAI large-v3-turbo model
  • Speaker diarization: silero-vad + SpeechBrain ECAPA-TDNN with cosine clustering — no HuggingFace token required
  • Speaker rename: per-file modal with 10-second audio loop preview
  • Term correction: spaCy de_core_news_sm extracts named entities and proper nouns (e.g., place names, people, project names) sorted by frequency, editable in a modal — all occurrences replaced word-precisely in VTT/TXT/CSV
  • Pipeline: per-segment cuts at speaker boundaries → each whisper call sees exactly one speaker, word-accurate by construction (~99.8% frame-level agreement vs the previous pyannote reference on a 42-min benchmark audio)
  • Standalone CPython 3.13, all required dylibs, static ffmpeg

Requirements

  • macOS Apple Silicon (M1/M2/M3/M4) — arm64 only
  • macOS 11 (Big Sur) or newer
  • ~3 GB free disk

Bundle size: 1.8 GB compressed, 2.6 GB unpacked.

License

Application code is MIT (see LICENSE). The bundled Python runtime, models and binaries each come with their own licenses; the complete list with upstream links is in THIRD-PARTY-LICENSES.md. Required attributions for Apache 2.0 and LGPL components are in NOTICE.

Assets 3
Loading