README.md

metadata

license: cc-by-4.0
language:
  - or
  - en
name: ODEN‑speech
slug: oden-speech
categories: 100K<n<1M
source_datasets:
  - common_voice_17
  - ljspeech
  - libritts
  - vctk
  - indictts
  - mucs
  - sayantan_odia
custom pipeline_tag: automatic-speech-recognition

ODEN‑speech 🗣️🇮🇳

Odia Diverse ENsemble Speech Corpus

ODEN‑speech merges eight publicly‑available Odia (ଓଡ଼ିଆ) speech corpora into a single 16 kHz, speaker‑aware, text‑cleaned dataset suitable for ASR, TTS, representation learning and multilingual research.

---

✨ Highlights

🗂️ Source	Hours	License
Mozilla Common Voice 17 (Odia)	110 h	MPL‑2.0
LibriTTS (clean + other)	170 h	CC‑BY‑4.0
LJSpeech 1.1	24 h	CC‑BY‑4.0
VCTK (Odia & misc.)	40 h	CC‑BY‑4.0
IndicTTS (SPRING Lab)	35 h	CC‑BY‑SA‑4.0
MUCS 2023 (Odia)	50 h	CC‑BY‑SA‑4.0
Sayantan Odia TTS	18 h	CC‑BY‑SA‑4.0
Total	≈ 462 h	–

• Every WAV is re‑sampled to 16 kHz / mono.
• Text is normalised (Unicode NFC, punctuation cleanup) without losing Odia matras.
• Speaker / gender / duration / original‑dataset fields preserved.
• Stratified train / validation / test splits (90 / 5 / 5 %).

---

📦 Dataset structure

id:          string   # unique key
audio:       dict(path, bytes, sampling_rate)
text:        string   # normalised Odia sentence
speaker:     string   # e.g. cv_or_spk_42
gender:      string   # male | female | unknown
dataset:     string   # source corpus tag
duration:    float32  # seconds
sample_rate: int32    # 16000 for all

Tip: with datasets you can stream only the text column for language modelling:

ds = load_dataset('BBSRguy/ODEN-speech', split='train', streaming=True)
texts = ds.with_format("text")['text']

---

🚀 Usage

Automatic Speech Recognition (ASR)

from datasets import load_dataset, Audio

ds = load_dataset("BBSRguy/ODEN-speech", split="train", streaming=True)

def preprocess(batch):
    audio = batch["audio"]
    inputs = processor(audio["array"], sampling_rate=16_000, text=batch["text"])
    return inputs

asr_ds = ds.map(preprocess)

Text‑to‑Speech (TTS)

from datasets import load_dataset

ds = load_dataset("BBSRguy/ODEN-speech", split="train")
example = ds[0]
print(example["text"])
print(example["audio"]["path"])  # path on local cache

Inline audio preview

---

🏗️ Building the corpus

---

🔒 License

All constituent corpora are at least CC‑BY or CC‑BY‑SA. The merged dataset is distributed under CC‑BY‑4.0. Please credit “@BBSRguy · ODEN‑speech” in derivative works.

---

🙏 Acknowledgements

We thank Mozilla, SPRING Lab, CMU, the MUCS programme, and every volunteer contributor for making high‑quality Odia speech available.

---

👩‍💻 Contributing

Pull‑requests welcome! Upload additional Odia recordings (CC‑BY) or improved transcriptions and open an issue or PR.

---

Created with ❤️ by @BBSRguy – 2025‑05‑28