voidash/nepali-tts-ready

Nepali TTS-Ready Dataset

A curated, multi-source Nepali speech dataset prepared for text-to-speech (TTS) training.

Dataset overview

Property	Value
Total rows	515210
Total hours	1011.03
Dropped by SNR gate (≥ 9.0 dB)	12230
Languages	Nepali (primary), some code-mixed Nepali-English
Sample rate	16 kHz mono (materialized sources resampled as needed)

Sources

The dataset combines the following validated, single-speaker sources:

• TTS silver (YouTube/podcast candidates, diarized + listening reviewed)
• ASR base (clean ASR-promoted YouTube/podcast segments)
• Unique promotion (additional single-speaker diarized promotion segments)
• IndicVoices-R Nepali (public spontaneous/read speech)
• Public HF datasets: Bijay13 podcast-code-switch, amitpant7, biplov-thapa TTS-streamable, kshitizzzzzzz Nepali ASR, pujanpaudel, shunyalabs
• tonibirat/nepali-asr-benchmark-50gb (public ASR benchmark)
• Jenny038/nepali-training-data (public training data)
• OpenSLR43 Nepali female TTS (public read speech)
• Stop-go promising enhancement (hand-reviewed short segments)

Storage format

Audio is packaged in WebDataset-style tar shards under audio_shards/ to keep the repository size manageable and avoid Hub API rate limits. Each sample consists of a {key}.json metadata member and a {key}.bin audio member (raw 16 kHz mono WAV bytes).

Fields

• id: unique row identifier
• text: Nepali transcript
• duration_sec: audio duration
• audio: raw 16 kHz mono WAV bytes (decode with soundfile/librosa)
• source: source origin
• source_url: URL when available
• estimated_snr_db: estimated signal-to-noise ratio
• f0_mean_hz, f0_std_hz, voiced_ratio: fundamental frequency statistics
• emotion_top, emotion_top_prob: emotion predictions (TTS silver subset only)
• cer: character error rate when available from source ASR
• shard_idx, shard_key: location of the sample inside the tar shards

Usage

import json, io
import soundfile as sf
from datasets import load_dataset

# Streaming is recommended for a 160 GB dataset.
ds = load_dataset("webdataset", data_files="audio_shards/*.tar", split="train", streaming=True)

for sample in ds:
    meta = json.loads(sample["json"])
    audio_bytes = sample["bin"]
    audio_array, sample_rate = sf.read(io.BytesIO(audio_bytes), dtype="float32")
    print(meta["id"], meta["text"], sample_rate, len(audio_array))

License

Individual sources retain their original licenses. The dataset is a curated aggregation; verify source licenses before commercial use.