PleasedPenguin/naturalvoice_737h_48k

NaturalVoices — Sidon-Restored, UTMOS-Filtered (48 kHz)

High-quality English speech derived from NaturalVoices_VC_870h (JHU SmileLab), restored with Sidon v0.1 and kept only where restoration measurably improved perceptual quality (UTMOS gate). Each clip ships with rich per-utterance metadata (transcript, speaker age/gender, speaking rate, emotion, and quality scores) so it is ready for TTS / voice-cloning / ASR / paralinguistic research.

• 494,903 clips · 736.7 hours (clips ≥ 3.0 s)
• 48 kHz, mono, float32, embedded audio (plays in the HF viewer)
• 48 kHz twin: PleasedPenguin/naturalvoice_737h_16k — same utterances, different sampling rate

Quick start

from datasets import load_dataset

ds = load_dataset("PleasedPenguin/naturalvoice_737h_48k", split="train")
ex = ds[0]
print(ex["text"], ex["age"], ex["gender"], ex["wpm"], ex["utmos"])
audio = ex["audio"]            # {'array': np.ndarray, 'sampling_rate': 48000}

How it was built

1. Restore — every source clip is passed through Sidon v0.1 (deterministic feed-forward restoration, no sampling) to a 48 kHz waveform; the 16 kHz version is a clean downsample of that same restored signal.
2. Quality gate — UTMOS (utmos22_strong) is scored on the original and the restored clip. A clip is kept only if utmos > utmos_orig (restoration actually helped). Both scores are stored so you can re-threshold.
3. Label — speaker age (audeering wav2vec2-large-robust-24-ft-age-gender) and wpm / speaking rate (Qwen3 Forced Aligner 0.6B: aligned word count over the first-word-onset → last-word-offset span) are added. The original NaturalVoices annotations (emotion, SNR, DNSMOS, etc.) are carried through unchanged.

Columns

Column	Description
audio	Sidon-restored waveform (48 kHz, float32)
text	Transcript
file_name, Document, Part Number	Clip identifiers
total time, start, end	Clip duration / source offsets (seconds)
gender	Speaker gender (original NaturalVoices label)
age	Predicted speaker age in years (audeering wav2vec2)
wpm	Speaking rate (words per minute), forced-aligned
utmos	UTMOS of the restored clip
utmos_orig	UTMOS of the original clip (kept only when utmos > utmos_orig)
DNSMOSPro, SNR, ASR_CONF, NUM_SPKS	Original NaturalVoices quality annotations
Arousal, Dominance, Valence	Original emotion attributes
Neutral, Angry, Sad, Happy, Emotion	Original emotion labels

Dataset statistics

Speaking rate (wpm) — mean 175, median 171, p5–p95 104–261.

Age — mean 36.5 yr, median 33.7, p10–p90 22.8–55.4.

Gender — model vs. original-label agreement 94.2% (the original gender label is the one shipped in this dataset; the predictor was used only for cross-checking).

Source & license

Derived from NaturalVoices_VC_870h (JHU SmileLab). Audio has been restored/transformed; please honor the terms of the original NaturalVoices release for any redistribution or use. Restoration: Sidon v0.1. Quality metric: UTMOS (tarepan/SpeechMOS).

Citation

If you use this data, please cite the original NaturalVoices dataset:

@misc{du2025naturalvoiceslargescalespontaneousemotional,
      title={NaturalVoices: A Large-Scale, Spontaneous and Emotional Podcast Dataset for Voice Conversion},
      author={Zongyang Du and Shreeram Suresh Chandra and Ismail Rasim Ulgen and Aurosweta Mahapatra and Ali N. Salman and Carlos Busso and Berrak Sisman},
      year={2025},
      eprint={2511.00256},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2511.00256},
}