Enhancing Out-of-Vocabulary Performance of Indian TTS Systems for Practical Applications through Low-Effort Data Strategies
Paper • 2407.13435 • Published
Datasets:
Nepali OOV TTS Benchmark
Author: Anil Titung Tamang (Titung)
Based on: Anand et al. 2024 — IndicOOV (arxiv.org/abs/2407.13435)
GitHub: anil-titung-tamang/nepali-oov-tts-benchmark
Dataset Description
This dataset contains volunteer recordings and evaluation data for the Nepali OOV (Out-of-Vocabulary) TTS Benchmark — the first benchmark of its kind for Nepali TTS, adapted from the IndicOOV methodology.
What is OOV?
A word is OOV for a TTS model if it contains character bigrams unseen during training. These words are likely to be mispronounced.
Categories
| Code | Category | Examples |
|---|---|---|
| ABBR | Abbreviations | UNESCO, ATM, API |
| BRAND | Brand Names | Khalti, Daraz, Xiaomi |
| CM | Code-mixed | online buking, OTP pathaunu |
| CMPY | Company Names | Ncell, Worldlink |
| GOVT | Government Schemes | loksewa aayog |
| PROP | Proper Nouns | Dhaulagiri, Janakpur |
| NAV | Navigation | Baneshwor Chowk, Pulchowk |
Splits
train— 223 sentences, recorded by speaker_01 (44100 Hz WAV)
Evaluation Results (Step 7 — IER%)
Model: milanakdj/indic-parler-tts-nepali-finetuned-dgx-v2.0-cosine
| Category | OOV IER | IV IER |
|---|---|---|
| ABBR | 56.8% | 17.8% |
| BRAND | 21.6% | 18.7% |
| CM | 9.3% | 22.7% |
| CMPY | 17.8% | 21.9% |
| GOVT | 9.3% | 9.3% |
| PROP | 15.3% | 17.6% |
| NAV | 28.4% | 18.7% |
| Overall | 22.6% | 18.1% |
- Downloads last month
- 8