File size: 4,230 Bytes
673db71 601defb 673db71 fac15c3 673db71 864c690 673db71 fac15c3 673db71 fac15c3 673db71 fac15c3 673db71 864c690 fac15c3 673db71 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 | ---
license: cc-by-4.0
language:
- te
- hi
- ta
task_categories:
- text-to-speech
- audio-classification
tags:
- accent-evaluation
- phoneme-probe
- indic
- retroflex
- wav2vec2
- fad
- psd
size_categories:
- 1K<n<10K
viewer: false
---
# Praxel/psp-native-centroids
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](https://arxiv.org/abs/2604.25476)
(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](https://github.com/praxelhq/psp-eval)
to score TTS outputs on six accent dimensions.
## Contents
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](https://www.iitm.ac.in/donlab/tts/) (Telugu subset) — CC-BY-4.0
- **Hindi**: [Rasa](https://github.com/AI4Bharat/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](https://github.com/praxelhq/psp-eval); see that
script for the exact extraction pipeline and the alignment-model checkpoints used.
## Usage
```python
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:
```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
```bibtex
@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).
## Related
- **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
## Contact
Pushpak Teja — pushpak@praxel.in — [praxel.in](https://praxel.in)
|