Update dataset card

README.md

@@ -1,39 +1,19 @@
 ---
-dataset_info:
-  config_name: read
-  features:
-  - name: id
-    dtype: string
-  - name: audio
-    dtype:
-      audio:
-        sampling_rate: 48000
-  - name: text
-    dtype: string
-  - name: speaker_id
-    dtype: int32
-  - name: style
-    dtype: string
-  - name: substyle
-    dtype: string
-  - name: corpus
-    dtype: string
-  - name: start_s
-    dtype: float32
-  - name: end_s
-    dtype: float32
-  splits:
-  - name: train
-    num_bytes: 3481898579
-    num_examples: 10388
-  - name: dev
-    num_bytes: 200973683
-    num_examples: 628
-  - name: test
-    num_bytes: 192517925
-    num_examples: 588
-  download_size: 3712884404
-  dataset_size: 3875390187
+license: cc-by-nc-4.0
+language:
+- en
+task_categories:
+- text-to-speech
+- automatic-speech-recognition
+tags:
+- expressive-speech
+- expresso
+- emotional-speech
+- style-transfer
+- multi-speaker
+pretty_name: Expresso (audio + text)
+size_categories:
+- 10K<n<100K
 configs:
 - config_name: read
   data_files:
@@ -44,3 +24,119 @@ configs:
   - split: test
     path: read/test-*
 ---
+
+# Expresso — audio + text
+
+A faithful re-publication of the official [Expresso](https://speechbot.github.io/expresso/)
+dataset (Nguyen et al., Interspeech 2023) as a loadable HuggingFace audio dataset, sourced
+directly from FAIR's official tar.
+
+> ⚠️ **License: CC-BY-NC-4.0** — non-commercial use only.
+
+## Status
+
+- ✅ **`read`** — published below (this card).
+- ⏳ `conversational` — coming soon (per-channel VAD-segmented mono utterances with machine transcripts).
+
+## `read` config
+
+11.6k mono utterances at **48 kHz / 24-bit**, fully transcribed by humans.
+
+| | train | dev | test |
+|---|---|---|---|
+| rows | 10,388 | 628 | 588 |
+
+### Schema
+
+| Column | Type | Notes |
+|---|---|---|
+| `id` | string | e.g. `ex01_confused_00001`; longform chunks: `ex01_default_longform_00001__0-16.49` |
+| `audio` | Audio @ 48 kHz mono | |
+| `text` | string | human-written transcription (mixed case, with punctuation) |
+| `speaker_id` | int32 | 1–4 |
+| `style` | string | one of: `default`, `confused`, `enunciated`, `happy`, `laughing`, `narration`, `sad`, `whisper` |
+| `substyle` | string | finer-grained label, e.g. `default_emphasis`, `default_essentials`, `default_longform`, `narration_longform` |
+| `corpus` | string | `base` (short utterances) or `longform` (multi-minute readings) |
+| `start_s` | float32 | null for full-file rows; chunk start for longform |
+| `end_s` | float32 | null for full-file rows; chunk end   for longform |
+
+### Splits
+
+We follow the official Expresso train/dev/test splits, with **one TTS-oriented deviation**:
+
+- **base read** (~11,600 utterances): full-file rows, no slicing — official splits applied as-is.
+- **longform read** (8 source files: `default_longform`, `narration_longform` × 4 speakers): kept as **full files in `train` only**. The official Expresso splits slice each longform file into 3 non-overlapping chunks (60 s for dev/test, the rest for train) for resynthesis benchmarking. Those chunks don't align with the full-file transcripts, so for TTS/ASR we keep the longform audio + transcript intact and place the full files in `train` only. If you need the official chunked benchmark, see `original_metadata/splits/`.
+- **singing** is intentionally **excluded** (only 12 wavs total, not in official splits).
+
+All rows have aligned `(audio, text)` pairs.
+
+### Style coverage per speaker
+
+All 4 speakers have all 8 styles, with these caveats:
+- `narration` is **longform-only** for all speakers (1 file each).
+- `default` includes the substyles `default`, `default_emphasis`, `default_essentials`, `default_longform`.
+
+## Sidecar files
+
+The original FAIR metadata is uploaded under `original_metadata/`:
+- `original_metadata/README.txt`, `LICENSE.txt` — official Expresso documentation
+- `original_metadata/read_transcriptions.txt` — per-file transcripts (tab-separated)
+- `original_metadata/VAD_segments.txt` — per-channel VAD timings for the conversational subset (used to derive the `conversational` config)
+- `original_metadata/splits/{train,dev,test}.txt`, `splits/README` — official split definitions
+
+## Quick start
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("shangeth/expresso", "read", split="train")
+ex = ds[0]
+print(ex["id"], "|", ex["style"], "|", ex["text"])
+print(ex["audio"]["array"].shape, "@", ex["audio"]["sampling_rate"], "Hz")
+
+# Filter to a specific style
+whisper = ds.filter(lambda x: x["style"] == "whisper")
+print(len(whisper), "whispered utterances")
+
+# Per-speaker breakdown
+from collections import Counter
+print(Counter(ds["speaker_id"]))
+```
+
+## Reproducing this dataset
+
+```bash
+# Download the official Expresso tar (~36 GB) and extract:
+mkdir -p data && cd data
+curl -L https://dl.fbaipublicfiles.com/textless_nlp/expresso/data/expresso.tar | tar -xf -
+cd ..
+
+# Build + push:
+python expresso_audio.py --repo_id shangeth/expresso --private
+```
+
+See [github.com/shangeth/wren-datasets](https://github.com/shangeth/wren-datasets) for the full extraction code.
+
+## Citation
+
+```bibtex
+@inproceedings{nguyen2023expresso,
+  title     = {Expresso: A Benchmark and Analysis of Discrete Expressive Speech Resynthesis},
+  author    = {Nguyen, Tu Anh and Hsu, Wei-Ning and D'Avirro, Antony and Shi, Bowen and
+               Gat, Itai and Fazel-Zarani, Maryam and Remez, Tal and Copet, Jade and
+               Synnaeve, Gabriel and Hassid, Michael and Kreuk, Felix and Adi, Yossi and Dupoux, Emmanuel},
+  booktitle = {Interspeech},
+  year      = {2023}
+}
+
+@misc{wren2026,
+  title  = {Wren: A Family of Small Open-Weight Models for Unified Speech-Text Modelling},
+  author = {Shangeth Rajaa},
+  year   = {2026},
+  url    = {https://github.com/shangeth/wren}
+}
+```
+
+## License
+
+**CC-BY-NC-4.0** — non-commercial use only. See `original_metadata/LICENSE.txt`.