Native-speaker reference artefacts for the PSP (Phoneme Substitution Profile) benchmark for Indic text-to-speech accent evaluation. Companion to the paper PSP: An Interpretable Per-Dimension Accent Benchmark for Indic Text-to-Speech (Menta, 2026).

This dataset is a scoring reference, not a training corpus. It contains pre-computed acoustic references extracted from publicly-licensed native-speaker speech corpora, used by the psp-eval package to score TTS outputs on six accent dimensions.

Per-language files for Telugu (te), Hindi (hi), and Tamil (ta):

File	Shape / size	Description
`{lang}_refs.pkl`	`{phoneme: [ndarray (1024,)]}`	Per-phoneme Wav2Vec2-XLS-R layer-9 centroid bags (500-clip bootstrap)
`{lang}_fad_natives.pkl`	`ndarray (1000, 1024)`	Utterance-level XLS-R embeddings for FAD computation
`{lang}_psd_natives.pkl`	`ndarray (500, 5)`	Prosodic feature vectors (F0 mean/std/range, onset-rate, nPVI) for PSD
`{lang}_sanity.json`	small JSON	Held-out native-audio sanity-check scores (§6 paper Signal 5)

Provenance

All centroids and reference distributions are derived from:

Telugu: IndicTTS (Telugu subset) — CC-BY-4.0
Hindi: Rasa (Hindi subset) — CC-BY-4.0
Tamil: IndicTTS (Tamil subset) — CC-BY-4.0

500 clips per language sampled from the full corpus with seed 1337. FAD references sample 1000 clips from the same pool with the same seed. PSD references sample 500 clips. Held-out sanity-check clips sample from the same pool with disjoint seed 999.

Each pickle was produced by evaluation/psp_bootstrap.py in the praxelhq/psp-eval repository; see that script for the exact extraction pipeline and the alignment-model checkpoints used.

Usage

from psp_eval import score_directory

# Centroids auto-download from this repo on first use.
scores = score_directory("my_tts_outputs/", language="te")

Or load directly in Python:

import pickle
from huggingface_hub import hf_hub_download

path = hf_hub_download("Praxel/psp-native-centroids", "te_refs.pkl", repo_type="dataset")
with open(path, "rb") as f:
    refs = pickle.load(f)
# refs: {"ṭ": [np.ndarray (1024,), ...], "ḍ": [...], ...}

Known caveats

Per-phoneme probe noise floor: native Telugu / Tamil audio registers 0.47–0.54 retroflex fidelity when scored against these centroids (not 1.0). This reflects speaker variance between centroid and held-out native corpora, aligner quality, and the strictness of the 0.5 collapse threshold. Interpret per-phoneme scores as relative rankings across systems, not absolute distances from a theoretical 1.0 ceiling. See paper §6 Signal 5 for details.
FAD / PSD do not share this noise floor (native audio correctly scores 5–50× lower than commercial-TTS outputs).
Unnormalised Fréchet across mixed-scale PSD dimensions: nPVI has numeric range ~$10^2$ while log-$F_0$ is ~$10^0$. A z-scored variant is planned for the v2 release.

Citation

@misc{teja2026psp,
  title={{PSP}: An Interpretable Per-Dimension Accent Benchmark for Indic Text-to-Speech},
  author={Menta, Venkata Pushpak Teja},
  year={2026},
  eprint={2604.25476},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2604.25476}
}

License

CC-BY-4.0 — matching the originating corpus licenses (IndicTTS, Rasa).

Code: https://github.com/praxelhq/psp-eval (MIT)
Install (from source): pip install git+https://github.com/praxelhq/psp-eval.git (PyPI publish planned post-paper-acceptance)
Paper: https://arxiv.org/abs/2604.25476