blizzard-indian-g2p-dur-eval
DNN-based Speech Synthesis for Indian Languages from ASCII text — Ronanki et al. (2016) (arXiv:1608.05374, 2016)
What this evaluates
Evaluates DNN-based grapheme-to-phoneme and duration prediction models for three Indian languages (Hindi, Tamil, Telugu) using crowdsourced ASCII transliterated text. It measures the accuracy of predicted phoneme durations against forced-aligned ground truth to assess component quality for speech synthesis.
Datasets
- Blizzard Challenge 2015 (Hindi, Tamil, Telugu) — total 5653; splits: train (-1), dev (-1), test (-1)
Metrics
RMSE (frames per phone)(primary) — range: other- Root Mean Square Error between predicted and forced-aligned phoneme durations. Calculated as the square root of the average of squared differences between predicted and ground truth duration values.
Input / output format
Input: ASCII transliterated text, tokenized by whitespace and punctuation.
Output: Predicted phoneme durations in frames.
Scoring recipe
def compute_rmse(predictions, gold):
if len(predictions) != len(gold):
raise ValueError('Length mismatch')
squared_diffs = [(p - g) ** 2 for p, g in zip(predictions, gold)]
mse = sum(squared_diffs) / len(squared_diffs)
return math.sqrt(mse)
Common pitfalls
- Splits are stratified by annotator rather than randomly split, meaning different annotators produced the train, dev, and test data to prevent leakage.
- The train/dev/test split ratios are 92%/4%/4%, which deviates from the common 90/5/5 or 80/10/10 conventions and may cause confusion.
Evidence (verbatim from paper)
Table 4: RMSE (frames per phone) between predicted and forced-aligned durations.
Citation
@misc{ronanki2016dnn,
title={DNN-based Speech Synthesis for Indian Languages from ASCII text},
author={Ronanki et al. (2016)},
year={2016},
note={arXiv:1608.05374}
}
- arXiv: 1608.05374