Local Transcriber 0.9

AI evaluation: turn conversations into minutes, action items and more – with a local model on your Mac, or a cloud model with names protected. Plus a glossary in the app and better recognition of difficult passages.

Install or update (Apple Silicon, macOS 14+): in the app, “Copy update command”, or

curl -fsSL https://transcriber.teconomix.net/install.sh | bash

Homebrew: brew upgrade --cask local-transcriber.

New: evaluate with AI (Settings › AI)

• Templates: minutes (key points, decisions, action items, open questions), short summary, order in the five-paragraph format (situation, mission, execution, support, responsibility & communication – with “still to clarify”), supervision and coaching note (without interpretation), key statements with quotes, follow-up email. Edit them, reset them, or add your own.
• Local or cloud, your choice: any server with an OpenAI-style API – locally Ollama, LM Studio, MTPLX or your own address, so the conversation never leaves the Mac; or OpenAI, Anthropic, Mistral, OpenRouter with your own key (stored in the keychain).
• Names stay with you: before a transcript goes to a cloud model, the names of speakers and calendar attendees (first names too) are replaced with placeholders and put back only in the finished summary on your Mac. On by default.
• Cloud only on click: cloud models never run automatically unless you explicitly allow it. Local models can summarize every new transcript automatically.
• In the panel: “Summarize” below a transcript; the entry shows the first key points and action items with Open · Copy · Share · ⋯. Share opens the macOS share menu – Mail gets a subject and formatted text. Continue in Claude, Claude Code or Codex opens the app with a prepared text (nothing is sent).
• The summary is saved as <name>.summary.md next to the transcript, moves along when you rename, and picks up speaker names you assign later.
• MCP: new tools recent_transcripts, read_transcript, get_summary, list_templates, summarize_transcript; every template is also a prompt (e.g. /protokoll in Claude Desktop) that your AI client applies with its own model.
• Obsidian plugin 0.2.0 (obsidian-local-transcriber): the summary appears as its own section above the transcript – checking off action items does not stop the transcript from updating – and a “Summarize” command.
• Measured on a MacBook Pro M5 Pro with Qwen3.8-27B: a 25-minute conversation in 30–50 s, three hours in about four minutes.

New: glossary and recognition

• Glossary in the app (Settings › Glossary): correct spellings of names and terms with the wrong spellings speech recognition produces; also inside hyphenated words. Apply it to existing transcripts with Re-analyze › Apply glossary. MCP: show_glossary, add_to_glossary, apply_glossary.
• Passages without punctuation are recognized again on their own and only replaced if the new text checks out; word loops are shortened.
• Unknown voices get a “probably” name less often (threshold 0.50 → 0.60).

Changed

• Settings with a sidebar (the tab bar no longer fit).
• Rename by clicking the title of an entry; the separate link is gone.
• The panel background is more opaque, so links stay readable on any desktop picture.

Fixed

• Renaming that only changed upper/lower case added “ 2” to the name (macOS does not distinguish case in file names).
• A transcript was not recognized as Transcriber's own when the file name came from the phone in a different Unicode form, so a second transcript file appeared next to it.