transkription

---

name: transkription version: 0.1.0 type: assist author: ellmos-ai created: 2026-06-22 updated: 2026-06-22 description: Transcribes audio/video files to text. Uses Whisper (openai-whisper) or Vosk (offline) as optional backend — both are detected via presence check. Without backend: placeholder mode with dummy output (dry-run).

standalone: true anthropic_compatible: true bach_compatible: false bach_origin: false category: assist tags: [transkription, audio, speech-to-text, whisper, vosk, offline] language: de status: stable dependencies: {'tools': [], 'services': [], 'protocols': [], 'python': [{'name': 'openai-whisper', 'optional': True, 'install': 'pip install openai-whisper', 'purpose': 'STT backend option 1 (cloud/local model)'}, {'name': 'vosk', 'optional': True, 'install': 'pip install vosk', 'purpose': 'STT backend option 2 (fully offline)'}]} provenance: {'origin': 'eigenentwurf', 'origin_path': '', 'origin_version': '', 'origin_repo': '', 'origin_license': 'MIT', 'last_sync_from_origin': '', 'notes': 'Kein direkter BACH-Origin vorhanden (transkriptions-service existiert nicht als Datei in BACH/system). Skill neu konzipiert. voice_stt.py aus BACH/hub/_services/voice/ hat das Backend-Muster inspiriert (optionale Imports mit Verfügbarkeits-Flags), wurde aber nicht direkt portiert.\n'}

Deutsch — Offizielle Deutsch-Version / Documento Oficial en Deutsch.

Übersicht & Zweck

Convert audio/video files to text — locally, without mandatory cloud access. The skill automatically detects whether Whisper or Vosk is installed and selects the best available backend. Without a backend it runs in dry-run mode and returns a placeholder text, so the workflow always works.

Transcripts are stored locally in transkription/store.db and can be queried.

---

Triggers

Phrase	Action
"Transcribe this audio"	Transcribe audio file
"Transcribe [file]"	Transcribe named file
"Show my transcripts"	List latest transcripts
"Search transcript [term]"	Full-text search in transcripts
"Export transcript [ID]"	Export transcript as TXT

---

Workflow & Vorgehen

1. Backend check: Check whether whisper or vosk is importable.
2. File check: Input file must exist (audio: wav, mp3, m4a, ogg, flac; video: mp4, mkv, webm — extraction via ffmpeg).
3. Transcription: Call backend and obtain raw text.
4. Save: Store result with metadata (file, duration, language, backend, timestamp) in store.db.
5. Output: Return text; optionally export as .txt.

---

CLI Entry Point

# Transcribe file (Deutsch)
python transkription_core.py transcribe audio.wav

# With explicit language (Deutsch)
python transkription_core.py transcribe audio.mp3 --lang de

# Dry-run (no backend required) (Deutsch)
python transkription_core.py transcribe audio.wav --dry-run

# List transcripts (Deutsch)
python transkription_core.py list [--limit 20]

# Full-text search (Deutsch)
python transkription_core.py search "term"

# Export (Deutsch)
python transkription_core.py export <id> [--out file.txt]

# Backend check (Deutsch)
python transkription_core.py check

# Alternative store path (e.g. for tests) (Deutsch)
python transkription_core.py --store /tmp/test.db transcribe audio.wav --dry-run

---

Store

Property	Value
Type	SQLite
Path (default)	skills/assist/transkription/store.db
Override	--store <path> or env TRANSKRIPTION_STORE
Tables	transcripts

Schema transcripts

CREATE TABLE IF NOT EXISTS transcripts (
    id          TEXT PRIMARY KEY,  -- UUID (short: 8 hex)
    file_path   TEXT NOT NULL,     -- original path of audio file
    file_name   TEXT NOT NULL,     -- filename (without path, for display)
    text        TEXT NOT NULL,     -- transcribed text
    language    TEXT,              -- language (e.g. "de", "en")
    backend     TEXT,              -- "whisper" | "vosk" | "dry-run"
    duration_s  REAL,              -- duration in seconds (if known)
    created_at  TEXT NOT NULL,     -- ISO-8601 timestamp
    tags        TEXT               -- comma-separated tags (optional)
);

---

Attitude

• Without an installed backend the skill works in dry-run mode (demo text).
• Whisper is preferred over Vosk (better German quality).
• The choice between Whisper and Vosk can be set via assist/prefs.json (transkription_backend: "whisper"|"vosk"|"auto").
• ffmpeg for video extraction is needed separately and is not included in the skill.

---

Privacy

• All transcripts stay local — no cloud transfer without Whisper online mode.
• Whisper can be used locally (tiny/base/medium model) or via OpenAI API. By default the local model is used.
• store.db may contain sensitive conversation content — do not commit to Git.
• Recommendation: add store.db to .gitignore.

---

Related Resources

• BACH hub/_services/voice/voice_stt.py — backend pattern (inspiration, read-only)
• Skill utilities/yt-transcriber — YouTube transcription (separate skill, not a duplicate: YT-specific)
• tools/module-installer/module_installer.py — registry contains whisper + vosk

---

Änderungsprotokoll

Version	Date	Change
0.1.0	2026-06-22	Initial creation — own SQLite store, Whisper/Vosk presence check