multi_accent_speech / README.md
cagatayn's picture
Update README.md
2038a7a verified
# Multi-Accent English Speech Corpus (Augmented & Speaker-Disjoint)
This dataset is a curated and augmented multi-accent English speech corpus designed for **speech recognition, accent classification, and representation learning**.
It consolidates multiple open-source accent corpora, converts all audio to a unified format, applies targeted data augmentation, and exports in a tidy, Hugging Face–ready structure.
---
## ✨ Key Features
- **Accents covered (12 total):**
american_english, british_english, indian_english, canadian_english, australian_english, scottish_english, irish_english, new_zealand_english, northernirish, african_english, welsh_english, south_african_english
- **Speaker-disjoint splits**: each speaker is assigned to exactly one split (train/validation/test).
- **Augmentation strategy**:
- < 2.6k samples → expanded to 5k via augmentation
- 2.6k–10k samples → expanded to 10k via augmentation
- > 10k samples → 50% replaced in place with augmented versions
- Aug methods: time-stretch, pitch-shift, background noise injection, reverb/EQ
- **Audio format (standardized)**: `.wav`, 16-bit PCM, 16 kHz sample rate, mono
- **Metadata-rich**: `uid`, `text`, `accent`, `speaker_id`, `dataset`, `is_augmented`, `source_uid`, `aug_label`, `duration_s`
---
## 📂 Dataset Structure
```
hf_export/
├── data/
│ ├── train/
│ │ ├── 0000/uid.wav
│ │ └── ...
│ ├── validation/
│ └── test/
├── metadata/
│ ├── train.parquet
│ ├── validation.parquet
│ └── test.parquet
├── SPLIT_REPORT.md
├── QA_REPORT.md
├── splits_by_speaker.csv
└── README.md
```
---
## 🗂 Metadata Schema
| Column | Type | Description |
|--------------|----------|-------------|
| `uid` | string | Unique ID per sample |
| `path` | string | Relative path to the `.wav` file |
| `text` | string | Transcript |
| `accent` | string | One of 17 whitelisted accents |
| `speaker_id` | string | Unique speaker ID (dataset-prefixed) |
| `dataset` | string | Source dataset (cv, vctk, accentdb, etc.) |
| `is_augmented` | bool | True if augmented |
| `source_uid` | string | UID of the original sample (for augmented rows) |
| `aug_label` | string | Applied augmentation method (e.g., `pitch:+1.2`) |
| `duration_s` | float32 | Duration in seconds |
Note: for accents with over 10k samples, is_augmented is false even though there's augmentation for half of their size. Use random sampling for these accents.
---
## 📊 Splitting Strategy
- **Ratios**: 78% train, 11% validation, 11% test
- **Speaker-disjoint**: a speaker appears in only one split
- **Accent-aware**: splits preserve global accent proportions
- **Dataset balance**: allocator prefers splits underrepresented for a dataset
- **Augmentation inheritance**: augmented samples inherit the split of their source speaker
---
## ✅ Preflight Validations
Checks applied before release:
- Unique `uid` values
- All files exist and are readable
- Format: 16kHz / mono / PCM_16
- `accent` ∈ whitelist (17)
- Transcripts non-empty
- Augmented samples link back to valid originals
- Duration bounds: 0.2s–30s (flagged outliers)
- Speaker-disjointness across splits
- Accent & dataset distributions close to global ratios
- Duplicate detection (duration+text fingerprint)
Reports:
- **SPLIT_REPORT.md**: speaker allocation, accent/dataset balance
- **QA_REPORT.md**: split sizes, duration anomalies, duplicate candidates
---
## 🔎 Example Usage
```python
from datasets import load_dataset
ds = load_dataset("cagatayn/multi_accent_speech", split="train")
print(ds[0])
# {
# 'uid': 'abc123',
# 'path': 'data/train/0000/abc123.wav',
# 'text': "it's the philosophy that guarantees everyone's freedom",
# 'accent': 'american_english',
# 'speaker_id': 'cv_sample-000031',
# 'dataset': 'cv',
# 'is_augmented': False,
# 'source_uid': '',
# 'aug_label': '',
# 'duration_s': 3.21,
# 'audio': {'path': 'data/train/0000/abc123.wav', 'array': ..., 'sampling_rate': 16000}
# }
```
---
## 📜 License
- Audio/data sourced from **Common Voice, VCTK, L2 Arctic, Speech Accent Archive, AccentDB**.
- Augmented versions generated as derivative works.
- Redistribution follows the most restrictive upstream license. Please review original dataset terms before commercial use.
---
## 🙏 Acknowledgments
- Mozilla [Common Voice](https://commonvoice.mozilla.org/)
- [VCTK Corpus](https://datashare.ed.ac.uk/handle/10283/3443)
- [Speech Accent Archive](https://accent.gmu.edu/)
- [AccentDB](https://github.com/AI4Bharat/AccentDB)
- [L2-ARCTIC Corpus](https://psi.engr.tamu.edu/l2-arctic-corpus/)
---