Update README.md

README.md

@@ -22,4 +22,71 @@ configs:
   data_files:
   - split: train
     path: data/train-*
+task_categories:
+- automatic-speech-recognition
+language:
+- de
+pretty_name: S
 ---
+
+# Dataset Card: Swiss German ↔ Standard German Speech Corpus (SPC) — Train v0.9
+
+## Summary
+
+The SPC Train v0.9 release pairs **Swiss German speech** with **Standard German transcriptions**, providing a high‑quality resource for training and evaluating automatic speech‑recognition (ASR) or speech‑translation systems.
+If you intend to fine‑tune Whisper, we recommend the companion project [`i4Ds/whisper‑finetune`](https://github.com/i4Ds/whisper-finetune), which is fully compatible with the data structure produced here.
+
+---
+
+## Dataset Details
+
+### Generation Pipeline
+
+The corpus was created with [`i4Ds/whisper‑prep`](https://github.com/i4Ds/whisper-prep) using the following configuration:
+
+```yaml
+# Generation configuration
+maintain_speaker_chance: 0.50  # Probability of keeping the current speaker for consecutive utterances
+n_samples_per_srt: 120         # Number of audio fragments merged into each SRT file
+normalize_text: true           # Clean text according to rules in whisper_prep/generation/text_normalizer.py
+
+# Overlap settings
+# Overlaps are inserted only in non‑speech regions identified by VAD.
+overlap_chance: 0.80           # Probability of creating an overlap between consecutive clips
+max_overlap_chance: 0.50       # If an overlap occurs, probability of using the maximum duration
+max_overlap_duration: 0.30     # Maximum overlap length in seconds
+```
+
+**Parameter glossary**
+
+| Parameter                 | Description                                                                                             |
+| ------------------------- | ------------------------------------------------------------------------------------------------------- |
+| `maintain_speaker_chance` | Likelihood that adjacent utterances originate from the same speaker, enabling more natural dialog flow. |
+| `n_samples_per_srt`       | Target number of utterances combined into a single subtitle (SRT) segment.                              |
+| `normalize_text`          | Applies rule‑based cleanup (e.g., punctuation, casing) to canonicalise transcripts.                     |
+| `overlap_chance`          | Chance of introducing slight temporal overlaps to mimic conversational turn‑taking.                     |
+| `max_overlap_chance`      | When an overlap is triggered, probability of shortening silence completely (back‑to‑back speech).       |
+| `max_overlap_duration`    | Hard cap on overlap length, preventing excessive speech collision.                                      |
+
+### Maintainer
+
+* **Curated by:** [Vincenzo Timmel](mailto:vincenzo.timmel@fhnw.ch) (@vincenzo.timmel)
+
+---
+
+## Intended Use & Scope
+
+* **Primary use‑case:** Fine‑tuning and evaluating multilingual ASR or speech‑translation models, particularly OpenAI Whisper.
+* **Not suitable for:** Language‑identification or emotion‑recognition tasks without additional annotation.
+
+---
+
+## Dataset Sources
+
+* **Related papers:** [“Swiss Parliaments Corpus”](https://arxiv.org/pdf/2010.02810), ["Fine-tuning Whisper on Low-Resource Languages"](https://arxiv.org/abs/2412.15726)
+
+---
+
+## Citation
+
+If you use this corpus, please cite the paper above and acknowledge **I4DS FHNW** for data preparation.