raw-emocean

Large-scale English speech dataset for text-to-speech (TTS) model training. Designed for autoregressive TTS architectures (TADA, CSM, VALL-E style models).

Dataset Summary

Metric	Value
Parquet shards	7
Segment duration	3–8 seconds
Sample rate	24,000 Hz (mono)
ASR engine	NVIDIA Parakeet TDT 0.6B v3
Format	Parquet with embedded audio

Dataset Schema

Column	Type	Description
audio	Audio	Waveform array + sampling rate (24kHz mono)
text	string	ASR transcript (Parakeet TDT 0.6B v3, no normalization)
duration	float64	Segment duration in seconds
snr	float64	Estimated signal-to-noise ratio (dB)
video_id	string	YouTube video identifier
source_url	string	Full source URL
start_time	float64	Segment start offset in source audio (seconds)
end_time	float64	Segment end offset in source audio (seconds)

Quick Start

from datasets import load_dataset

# Stream large dataset without downloading entirely
ds = load_dataset("somu9/raw-emocean", split="train", streaming=True)
for sample in ds:
    print(sample["text"])
    break

# Load full dataset
ds = load_dataset("somu9/raw-emocean", split="train")
sample = ds[0]
audio_array = sample["audio"]["array"]       # numpy array
sample_rate = sample["audio"]["sampling_rate"]  # 24000
transcript = sample["text"]

# Filter by duration or SNR
clean = ds.filter(lambda x: x["snr"] > 30 and x["duration"] > 4)

Data Collection Pipeline

YouTube Audio
    │
    ▼
yt-dlp download (WAV 24kHz)
    │
    ▼
Voice Activity Detection (Silero VAD)
    │  threshold=0.5, min_speech=500ms, min_silence=300ms
    ▼
Segment Extraction (3–8s segments)
    │
    ▼
Quality Filters
    │  ├─ SNR > 25 dB
    │  ├─ Clipping < 0.1%
    │  ├─ Speaker overlap detection (pitch bimodality)
    │  └─ Music detection (spectral flatness + flux)
    │
    ▼
ASR Transcription (Parakeet TDT 0.6B v3)
    │  batch_size=128, CUDA
    ▼
manifest.csv + audio/*.wav

Quality Assurance

• SNR filtering: Segments below 25 dB signal-to-noise ratio are discarded
• Clipping detection: Segments with >0.1% of samples at peak amplitude are removed
• Speaker overlap: Pitch-based bimodality detection removes segments with simultaneous speakers
• Music detection: Spectral flatness and flux analysis removes segments with background music
• Minimum length: Segments shorter than 3 seconds are discarded
• Empty transcript filter: Segments where ASR produces fewer than 5 characters are removed

Intended Use

• Pre-training autoregressive TTS models (TADA, CSM, VALL-E, SoundStorm)
• Fine-tuning speech synthesis models
• Speech representation learning
• ASR training data augmentation

Limitations

• Transcripts are ASR-generated (Parakeet) — not human-verified, expect ~5% error rate
• Audio sourced from YouTube — variable recording conditions across sources
• No speaker identity labels — segments are not diarized
• No emotion or prosody annotations
• English only

Citation

If you use this dataset, please cite:

@dataset{raw-emocean_2026,
  title = {raw-emocean: English Speech Dataset for TTS Training},
  author = {Dataset Contributors},
  year = {2026},
  url = {https://huggingface.co/datasets/somu9/raw-emocean}
}

License

CC-BY-4.0

---

Last updated: 2026-04-25 14:27 UTC