DEUTSCH LERNEN — AUDIO SOURCES AND USE
Updated: 2026-10-26

The 32 B1 recordings and 13 weekend recordings in this version are synthetic
speech generated using the OpenAI API (gpt-realtime-2.1-mini). The voices
are built-in synthetic voices: marin, cedar, coral and sage. No human
speaker performed these exercises for the portal. No real person's voice
was cloned for this project. These are not official Goethe recordings.

Exercise text and scripts are independently created, AI-assisted portal
materials. Oliver Laudan is responsible for project/publication decisions.
The API receives the public exercise scripts during generation, not learner
profiles or user recordings. Portal playback serves static MP3 files and
does not call OpenAI.

Provenance:
- openai-generation.json: input text, voice assignments, API-returned
  transcripts, usage, model name and file hashes.
- audio-inventory.json: inventory of all 45 MP3 files.
- https://github.com/oliverlaudan-ops/learning-german/blob/main/docs/content-provenance.md

Provider terms:
https://openai.com/policies/services-agreement/
https://openai.com/policies/service-terms/
API documentation:
https://developers.openai.com/api/docs/guides/realtime-conversations

The repository's MIT license covers appropriately licensable project code;
it does not license OpenAI's models or voices themselves, override provider
terms, or guarantee copyright protection for purely machine-generated output.
These recordings were made through the API, not by capturing ChatGPT's
Read Aloud button or Voice conversations.

Earlier versions contained Piper-generated audio. Those MP3s and the legacy
Piper generation path are replaced in this version. Earlier source reviews
remain in Git history; replacing the recordings does not retroactively
determine the rights status of historical versions.

Quality:
Two initial samples passed Oliver's listening review. API-returned transcripts
are compared with input text, but this is not an independent hearing review.
Please review pronunciation, intelligibility and speaker distinction in the
complete exercises; errors remain possible.