RikkaBotan/nyan-jenny-format

Nyan Jenny Format

A Japanese text-to-speech (TTS) dataset in Jenny TTS format, derived from the multilingual_toxicity_dataset (Japanese split). Toxic words identified by DeepSeek V3 Pro are replaced with 「にゃん」(nyan) before audio synthesis, producing naturally-sounding speech that masks harmful content.

---

Dataset Summary

This dataset was created through the following pipeline:

1. Source data — The Japanese split of textdetox/multilingual_toxicity_dataset (5 k samples: 2.5 k toxic, 2.5 k non-toxic).
2. Toxic-word annotation — DeepSeek V3 Pro was used to identify toxic words and phrases within each toxic sample.
3. Substitution — Every annotated toxic word is replaced with「にゃん」, producing text that retains the original sentence structure while masking offensive content.
4. Speech synthesis — All texts (substituted toxic samples and original clean samples) are synthesized into audio with irodori TTS v3 600 M (voicedesign).
5. Formatting — The resulting audio and transcriptions are organized in the Jenny TTS column schema for compatibility with Jenny-based TTS training pipelines.

---

Dataset Details

Property	Value
Language	Japanese (ja)
Total rows	~9 810
Train split	~9 320 rows
Validation split	~491 rows
Audio format	WAV (mono)
Audio duration	1.6 – 30 seconds per sample
TTS model	Aratako/Irodori-TTS-600M-v3-VoiceDesign
Annotation model	DeepSeek V4 Pro
Format	Jenny TTS

---

Data Fields

Column	Type	Description
file_name	string	Unique identifier for each sample. Prefixed with toxic_ (from the toxic class, with にゃん substitutions applied) or clean_id_ (from the non-toxic class, unchanged). Length: 10–13 characters.
transcription	string	The text transcription of the audio. For toxic_ samples. Length: 11–139 characters.
transcription_normalised	string	Normalised version of the transcription. Currently identical to transcription.
audio	audio	Synthesized WAV audio corresponding to the transcription that toxic words have been replaced with「にゃん」.

---

Usage

from datasets import load_dataset

ds = load_dataset("RikkaBotan/nyan-jenny-format")

# Inspect a few samples
for sample in ds["train"].select(range(5)):
    print(sample["file_name"], "|", sample["transcription"])
    # sample["audio"] contains {"array": ..., "sampling_rate": ...}

Intended use

• Fine-tuning or training Japanese TTS models using the Jenny format.
• Research on speech synthesis with masked/detoxified text.
• Evaluation of TTS robustness with unusual token sequences (「にゃん」substitutions).

---

Source Dataset

textdetox/multilingual_toxicity_dataset

---

TTS Model

Audio was generated with Aratako/Irodori-TTS-600M-v3-VoiceDesign, a Japanese neural TTS system.

---

Limitations and Biases

• Residual context — While individual toxic words are replaced, the surrounding sentence context may still reflect the original tone or topic of toxic content.
• Synthesis artifacts — 「にゃん」substitutions may cause prosodic inconsistencies in synthesized speech, as the TTS model was not specifically trained on this token distribution.
• Single speaker — All audio is generated from a single TTS voice; speaker diversity is absent.
• Japanese only — Despite the multilingual source dataset, this release covers only the Japanese (ja) split.

---

License

This dataset inherits license constraints from the source data (openrail++) and the irodori TTS model. Please review the terms of both before redistribution or commercial use.