NepTTS-Bench Dataset

First comprehensive benchmark for evaluating Nepali text-to-speech systems.

Contents

• sentences.json — 365 phonologically-designed Nepali sentences with metadata
• audio/ — TTS outputs from 12 systems (~2000 audio files)
• results/ — Pre-computed evaluation results (SCOREQ, Chirp2, MMS, XLS-R, Whisper)
• baselines.json — Aggregate scores for all baseline systems
• model/ — NepaliMOS predictor checkpoint (Spearman 0.587)

Systems Evaluated

System	Human MOS	Type
Natural Speech	3.91	Human
TingTing Asmita	3.49	Nepali-specific
ElevenLabs v3	3.48	Cloud
Piper	3.47	Open source
TingTing Subina	3.42	Nepali-specific
Edge TTS Hemkala	3.31	Cloud (Microsoft)
Edge TTS Sagar	3.28	Cloud (Microsoft)
Gemini Flash	3.19	Cloud (Google)
gTTS	2.56	Cloud (Google Translate)

Usage

from datasets import load_dataset

ds = load_dataset("ampixa/neptts-bench")

Or use the evaluation CLI:

pip install neptts-eval
neptts-eval --wav_dir ./my_tts_outputs/

Citation

@article{neptts-bench-2026,
  title={NepTTS-Bench: A Comprehensive Benchmark for Nepali Text-to-Speech Evaluation},
  author={Ampixa},
  year={2026}
}

Links

• GitHub: Ampixa/neptts-bench
• Rate TTS: tts.ampixa.com/rating
• Paper: Coming soon