Update dataset card with conversational config

README.md

@@ -15,12 +15,6 @@ pretty_name: Expresso (audio + text)
 size_categories:
 - 10K<n<100K
 configs:
-- config_name: conversational
-  data_files:
-  - split: dev
-    path: conversational/dev-*
-  - split: test
-    path: conversational/test-*
 - config_name: read
   data_files:
   - split: train
@@ -29,42 +23,14 @@ configs:
     path: read/dev-*
   - split: test
     path: read/test-*
-dataset_info:
-  config_name: conversational
-  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: other_speaker_id
-    dtype: int32
-  - name: other_style
-    dtype: string
-  - name: source_file_id
-    dtype: string
-  - name: channel
-    dtype: int32
-  - name: start_s
-    dtype: float32
-  - name: end_s
-    dtype: float32
-  splits:
-  - name: dev
-    num_bytes: 287371683
-    num_examples: 521
-  - name: test
-    num_bytes: 293805348
-    num_examples: 515
-  download_size: 546394797
-  dataset_size: 581177031
+- config_name: conversational
+  data_files:
+  - split: train
+    path: conversational/train-*
+  - split: dev
+    path: conversational/dev-*
+  - split: test
+    path: conversational/test-*
 ---
 
 # Expresso — audio + text
@@ -75,10 +41,10 @@ directly from FAIR's official tar.
 
 > ⚠️ **License: CC-BY-NC-4.0** — non-commercial use only.
 
-## Status
+## Configs
 
-- ✅ **`read`** — published below (this card).
-- ⏳ `conversational` — coming soon (per-channel VAD-segmented mono utterances with machine transcripts).
+- **`read`** — 11.6k mono read-speech utterances with **human transcripts**.
+- **`conversational`** — ~15.9k mono per-utterance turns derived from the stereo conversational dialogues, transcribed with **Whisper Large V3 Turbo**.
 
 ## `read` config
 
@@ -118,6 +84,66 @@ 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`.
 
+---
+
+## `conversational` config
+
+~15.9k per-utterance mono turns derived from the official 339 stereo dialog files. Each row is **one speaker's turn** at a known time range within the source file, transcribed by Whisper.
+
+| | train | dev | test |
+|---|---|---|---|
+| rows | ~14.8k | ~520 | ~515 |
+| audio | ~29 h | ~50 min | ~51 min |
+
+### Schema
+
+| Column | Type | Notes |
+|---|---|---|
+| `id` | string | e.g. `ex01-ex02_default_001__ch1_23.88-28.14` |
+| `audio` | Audio @ 48 kHz mono | the VAD-extracted turn from one channel |
+| `text` | string | Whisper Large V3 Turbo transcript (mixed case + punctuation) |
+| `speaker_id` | int32 | this channel's speaker (1–4) |
+| `style` | string | this channel's expressive style |
+| `other_speaker_id` | int32 | partner's speaker id |
+| `other_style` | string | partner's expressive style |
+| `source_file_id` | string | e.g. `ex01-ex02_default_001` (the stereo source) |
+| `channel` | int32 | 1 or 2 |
+| `start_s` | float32 | turn start within source file (after VAD ∩ split clip) |
+| `end_s` | float32 | turn end |
+
+### How it was built
+
+1. **Parse** the official `splits/{train,dev,test}.txt` time-window assignments per source file.
+2. **Intersect** each split window with `VAD_segments.txt` (per-channel pyannote turns) — turns straddling the dev/test boundary are **clipped to the split window** so dev/test never leak into train.
+3. **Slice** the stereo source file → mono channel → 48 kHz mono turn.
+4. **Transcribe** with `openai/whisper-large-v3-turbo`, with anti-hallucination decoding (`no_repeat_ngram_size=4`, `repetition_penalty=1.2`, `condition_on_prev_tokens=False`) and pre-resampled to 16 kHz.
+
+### Turn filtering
+
+- **Min duration**: 0.3 s. Sub-300ms VAD turns (mostly backchannels and clicks) are dropped.
+- **Max duration**: 28 s. Long turns are split into ≤28 s pieces (Whisper's context is 30 s).
+
+### Style coverage
+
+26 styles total in the conversational subset, including styles **not present in `read`**: `angry`, `animal`, `awe`, `bored`, `calm`, `desire`, `disgusted`, `fast`, `fearful`, `nonverbal`, `projected`, `sarcastic`, `sleepy`, `sympathetic`, plus mixed pairs like `animal-animaldir` and `child-childdir` (where the two channels carry different styles — one row's `style` and `other_style` will differ).
+
+### ASR quality (validated against `read` ground truth)
+
+We benchmarked Whisper Large V3 Turbo on 210 human-transcribed read utterances spanning all 7 transcribed read styles. Per-style WER:
+
+| confused | default | sad | happy | enunciated | laughing | whisper | **overall** |
+|---|---|---|---|---|---|---|---|
+| 0.96% | 1.67% | 2.00% | 2.76% | 3.18% | 4.98% | 5.31% | **3.00%** |
+
+ASR errors are highest on `whisper` and `laughing` styles (the toughest acoustic conditions), but still under 6% WER. Conversational rows are expected to track the same per-style quality.
+
+### Caveats
+
+- Transcripts are **machine-generated** — expect a small error rate, especially on whispered/laughing/animal-style turns.
+- Mixed-style pairs (`animal-animaldir`, `child-childdir`, `sad-sympathetic` and reversals) — speakers in the two channels carry different styles. Ground-truth styles are encoded per-row in `style` (this channel) and `other_style` (partner).
+
+---
+
 ## Sidecar files
 
 The original FAIR metadata is uploaded under `original_metadata/`:
@@ -131,18 +157,21 @@ The original FAIR metadata is uploaded under `original_metadata/`:
 ```python
 from datasets import load_dataset
 
-ds = load_dataset("shangeth/expresso", "read", split="train")
-ex = ds[0]
+# Pick a config — there is no default
+read = load_dataset("shangeth/expresso", "read",          split="train")
+conv = load_dataset("shangeth/expresso", "conversational", split="train")
+
+ex = read[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")
+# Filter conv to mixed-style pairs (cross-style modeling)
+mixed = conv.filter(lambda x: x["style"] != x["other_style"])
+print(f"{len(mixed)} cross-style turns")
 
-# Per-speaker breakdown
+# Per-style coverage
 from collections import Counter
-print(Counter(ds["speaker_id"]))
+print(Counter(conv["style"]).most_common(10))
 ```
 
 ## Reproducing this dataset
@@ -154,7 +183,8 @@ curl -L https://dl.fbaipublicfiles.com/textless_nlp/expresso/data/expresso.tar |
 cd ..
 
 # Build + push:
-python expresso_audio.py --repo_id shangeth/expresso --private
+python expresso_audio.py        --repo_id shangeth/expresso --private  # read config
+python expresso_conversational.py --repo_id shangeth/expresso --private  # conversational config
 ```
 
 See [github.com/shangeth/wren-datasets](https://github.com/shangeth/wren-datasets) for the full extraction code.