audio-archive-tools

English · Deutsch · Français · Español · 日本語 · 中文

Turn a folder of audio recordings into a searchable, self-describing archive – locally, without a server and without uploading anything.

The tools transcribe recordings with noScribe (Whisper + speaker detection), create subtitles, give every file a date and a meaningful name, and build an offline overview page in which you can filter, search the full text and jump straight to any sentence in a recording. A form page lets you describe the project and each recording without touching a file by hand.

It started as a family archive of old voice-recorder and phone recordings and is designed to work just as well for:

Family / private	home recordings, children's voices, conversations with grandparents
Oral history	biographical and witness interviews with consent, access and embargo
Linguistic fieldwork	sessions with ISO 639-3 languages, speakers, genre, stimulus, consent and access level (IMDI/ELAN style)
Sociolinguistics	interviews with social variables and interview modules (conversation, reading passage, word list, minimal pairs)
Field recording / soundscape	nature and environmental sound with place, coordinates, habitat, equipment – no transcription
Journalism	takes and soundbites for reports, features, vox pops, demonstrations, press conferences – with release status

Features

• Batch transcription with noScribe in headless mode – restartable at any time, identical files are transcribed once, automatic second attempt without speaker detection, optional retry with audio clean-up for difficult recordings
• Subtitles (.srt) next to every recording – VLC shows them automatically; long passages are split into sentences
• Hallucination filter for typical Whisper artefacts in silence or noise (broadcaster credits, endless loops)
• Consistent file names: YYYY-MM-DD_Short-title_RecorderID.mp3 (estimated dates get ca), with an undo log
• Overview page: filters by year and by the archive's own tag categories, full-text search across all transcripts, built-in player, click a sentence to jump there, chapters or soundbites, project information
• Editor page: form for project details and every recording (date, title, summary, tags, profile fields, speaker names, chapters/soundbites) – changes are kept in the browser and saved as a file the tools import
• Profiles with field sets for the use cases above – freely adjustable per archive
• CSV export for spreadsheets, corpus tools or archive deposits
• Six languages for pages, messages and batch file names: English, German, French, Spanish, Japanese, Chinese
• Duplicates, AMR originals (converted to MP3 for browsers) and empty files are moved aside, never deleted
• Python standard library only – no packages to install

How it works

 recordings ──▶ 1 Transcribe ──▶ 2 Describe (editor page or Claude) ──▶ 3 Update archive ──▶ overview page
                 noScribe .html      archive.json                        rename, subtitles,     filter · search ·
                 + .srt                                                  build pages            listen · jump

Requirements

• Windows 10/11 (the batch files are for Windows; the Python scripts themselves are cross-platform)
• Python 3.9 or newer
• ffmpeg on the PATH (e.g. winget install ffmpeg or scoop install ffmpeg)
• noScribe 0.7 or newer (with command-line support). An NVIDIA GPU makes transcription much faster (about 10× real time on a GTX 1080 Ti), but is not required.

Installation

1. Download or clone this repository, e.g. to Documents\audio-archive-tools.

2. Copy settings.example.json to settings.json and adjust it:

   {
    "language": "en",
    "transcription_language": "auto",
    "noscribe_path": "C:/Program Files (x86)/noScribe/noScribe.exe",
    "model": "precise",
    "speaker_detection": "auto",
    "mark_pauses": "3sec+"
   }

   language is the language of pages, messages and batch files (en, de, fr, es, ja, zh).

Quick start

1. Back up the folder with your recordings.
2. Drag the folder onto setup-archive.bat. Choose a profile, a title and the language of the recordings. A subfolder _archive with the batch files appears.
3. Run 1 Transcribe. Every recording gets a .html transcript and a .srt subtitle file.
4. Run 2 Describe recordings (editor). Fill in the project details and describe the recordings, then save – the browser asks where to put the file: choose the _archive folder.
5. Run 3 Update archive. Files are renamed and the overview page Audio archive.html opens.

New recordings later? Put them into the folder and run 1 → 2 → 3 again. Only new files are processed.

Batch files in _archive

Batch file	What it does
1 Transcribe	transcribes everything that has no transcript yet and registers new recordings
2 Describe recordings (editor)	builds and opens the editor page
3 Update archive	imports saved edits, renames files, rebuilds subtitles and pages
Rebuild subtitles	after correcting a transcript in the noScribe editor
Re-transcribe a problem case	retry one recording with noise reduction and loudness normalisation
Export as CSV	writes <title>.csv and <title>_project.csv
Undo renaming	reverts all logged renames and moves

Profiles

When you set up an archive you choose a profile; its fields and tag categories are copied into _archive/archive.json, where you can rename, add or remove them. See docs/profiles.md for the fields of each profile and why they are there, and docs/data-format.md for the file format.

Describing recordings with Claude (optional)

Describing dozens of recordings is the part that takes time. With Claude Code you can let Claude read the transcripts and fill in dates, titles, summaries, tags, speaker names and chapters. Copy claude-skill/audio-archive to ~/.claude/skills/, adapt the two marked sections, and say for example "Describe the new recordings in my audio archive in D:\Fieldwork\2024". reading_view.py and apply_metadata.py are the tools the skill uses – they work just as well for any other workflow.

Privacy

Everything runs on your computer. noScribe transcribes locally, the pages are plain files that open from disk, nothing is uploaded. The .gitignore of this repository excludes audio files, _archive folders and your settings.json, so archive data cannot end up on GitHub by accident. If you use an AI assistant to describe recordings, the transcripts are shared with that service – decide per archive whether that is acceptable, particularly for research data with consent restrictions.

Limitations

• Automatic transcripts are drafts. Children's speech, dialects, noisy or very quiet recordings come out poorly. Speaker detection does not know names; assigning them is part of describing.
• Whisper does not cover every language. For under-resourced languages set "transcribe": false and use the archive for metadata, listening and chapter marks.
• No IMDI/CMDI XML export yet – the CSV export is the bridge to archive deposits.

Project structure

audio_archive/         Python tools (standard library only)
  locales/             en, de, fr, es, ja, zh
  profiles/            archive profiles
  template/            overview and editor pages
claude-skill/          optional Claude Code skill
docs/                  profiles and data format
tests/                 unit tests (python -m unittest discover -s tests)
setup-archive.bat      drag a folder onto it to set up an archive
settings.example.json  copy to settings.json

Contributing

See CONTRIBUTING.md – new languages and profiles are especially welcome.

License

MIT

Acknowledgements

Transcription is done by noScribe by Kai Dröge, which builds on faster-whisper and pyannote.