amiramini/german-learning-context
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Prefer official Hueber transcripts over ASR; ASR as fallback only

Browse Source
This commit is contained in:
amiramini
2026-08-28 21:35:35 +03:30
parent e65626cbc9
commit 3804ab5813
2 changed files with 21 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
A1.1/_template.md
+1 -2
View File
@@ -24,8 +24,7 @@ Source: [ ] Course book [ ] Workbook
## Audio / Listening ## Audio / Listening
Transcripts are ASR-generated (faster-whisper) unless noted otherwise — Transcript source: [ ] Official (Hueber Transkriptionen) [ ] ASR (faster-whisper — review for errors, especially names/numbers)
review for errors, especially names/numbers.
### Track [X.X] — [short description, e.g. "Anna and Ben introduce themselves"] ### Track [X.X] — [short description, e.g. "Anna and Ben introduce themselves"]
``` ```
README.md
+20 -7
View File
@@ -34,12 +34,24 @@ A1.2/ ← added once I move on to the next book
## Audio (listening exercises) ## Audio (listening exercises)
Course book audio has no printed transcript (the workbook's back section is Hueber (the publisher) releases official **Transkriptionen** PDFs for both
the *Lösungsschlüssel* — answer key — not a transcript, so that doesn't help the Kursbuch (course book) and Arbeitsbuch (workbook) — these are the
here). Raw audio files are **not** stored in this repo (see `.gitignore`); primary source for audio content, not the *Lösungsschlüssel* (answer key,
only their transcripts are, since that's what gives Claude context. found at the back of the workbook, which is a different thing). Raw audio
files are **not** stored in this repo (see `.gitignore`); only transcripts
(text) are, since that's what gives Claude context.
Workflow: Primary workflow — official transcripts:
1. Get the Kursbuch and Arbeitsbuch Transkriptionen PDFs for the relevant book
(A1.1, A1.2, ...) — search "Menschen [level] Kursbuch/Arbeitsbuch
Transkriptionen pdf Hueber", or check the publisher's site.
2. Send/upload the pages for a lesson the same way as course book/workbook
content — Claude extracts the per-track dialogue text into that lesson's
`.md` file under "Audio / Listening".
3. No review needed for accuracy — these are publisher-official text, not ASR.
Fallback workflow — local ASR (only if no official transcript exists for a
given track, e.g. supplementary listening material):
1. Get the audio files locally (publisher CD/app/download), any folder outside git. 1. Get the audio files locally (publisher CD/app/download), any folder outside git.
2. Run `scripts/transcribe.py` (uses `faster-whisper` on GPU) to transcribe them: 2. Run `scripts/transcribe.py` (uses `faster-whisper` on GPU) to transcribe them:
```bash ```bash
@@ -48,8 +60,9 @@ Workflow:
``` ```
3. Skim the `.txt` output for ASR mistakes (names, numbers, fast speech) and fix them. 3. Skim the `.txt` output for ASR mistakes (names, numbers, fast speech) and fix them.
4. Reference/paste the transcript into the lesson's `.md` file under "Audio / Listening", 4. Reference/paste the transcript into the lesson's `.md` file under "Audio / Listening",
or just point at the `transcripts/lektion-XX/*.txt` path — either works, transcripts flagged as ASR-generated (the template already does this), or point at the
are small text files so committing them is fine. `transcripts/lektion-XX/*.txt` path — transcripts are small text files so
committing them is fine.
## Adding a new book (e.g. A1.2) ## Adding a new book (e.g. A1.2)