README.md

---
license: other
license_name: see-per-model-licences
license_link: https://huggingface.co/AlphaAvatar/persona-speaker-attribute-onnx/blob/main/LICENSE
tags:
- onnx
- legacy
pipeline_tag: audio-classification
---

# AlphaAvatar Plugins Persona — LEGACY INDEX

> **This repository is legacy and is no longer the distribution point for the
> Persona ONNX models.** Branch-based model storage has been replaced by one
> repository per model. Nothing here has been deleted: every branch, artifact
> and commit is preserved exactly as it was.

## Where the models live now

| model | new repository | version tag | commit |
|---|---|---|---|
| speaker vector (ERes2NetV2, 192-d embedding) | [`AlphaAvatar/persona-speaker-vector-onnx`](https://huggingface.co/AlphaAvatar/persona-speaker-vector-onnx) | `v1.0.0` | `1899db09a40a60472681f07a189188517f515b4b` |
| speaker attribute (wav2vec2-large-robust 6L age/gender) | [`AlphaAvatar/persona-speaker-attribute-onnx`](https://huggingface.co/AlphaAvatar/persona-speaker-attribute-onnx) | `v1.0.0` | `3174195619187307495d5fa782a87dac86c5ddec` |

Pin the **commit**, not the tag or `main`.

## Legacy layout (preserved, do not use for new integrations)

| branch | commit | file | SHA256 (= Git-LFS oid) |
|---|---|---|---|
| `speaker_vector_onnx` | `8bb7633a3c0116ca68b0de476c65285c7820a1dd` | `eres2netv2.onnx` | `a8614dde1e71f5091ce35e031de672d4d87fe1dd839ffe968867febec67e8123` |
| `speaker_attribute_onnx` | `530618cc4ebd8c1aa2a995fc12a180d594535d3f` | `w2v2l6.onnx` | `75c5cc3debc2013215cee5f331a66b59bd7205da170fb61abab46fc4507df7be` |

The files in the new repositories are **byte-identical** to these; they were
renamed to `model.onnx` and nothing else changed. The SHA256 values above are
unchanged in the new repositories, so the migration is verifiable without
downloading both copies.

## Licence correction

The front matter of this repository previously declared `license: apache-2.0`
for both branches. That is correct for the speaker-vector model but **not** for
the speaker-attribute model:

* **speaker vector** — derived from ModelScope
  [`iic/speech_eres2netv2_sv_zh-cn_16k-common`](https://www.modelscope.cn/models/iic/speech_eres2netv2_sv_zh-cn_16k-common),
  **Apache-2.0**. Verified: three un-mangled ONNX initializers
  (`layer3_ds.weight`, `seg_1.weight`, `seg_1.bias`) are bit-identical to the
  upstream checkpoint.
* **speaker attribute** — derived from
  [`audeering/wav2vec2-large-robust-6-ft-age-gender`](https://huggingface.co/audeering/wav2vec2-large-robust-6-ft-age-gender),
  **CC-BY-NC-SA-4.0 (non-commercial)**. Verified: all 102 name-matched ONNX
  initializers are bit-identical to the upstream checkpoint, and the known-answer
  output published on the upstream model card is reproduced to 1.13e-06.

The new repositories carry the correct per-model licence, `LICENSE` and `NOTICE`.
See `AlphaAvatar/persona-speaker-attribute-onnx` for the non-commercial-use
consequences.

## Behavioural notes carried over to the new repositories

The speaker-attribute model has two behaviours that are easy to get wrong and are
documented in full in its new repository:

1. `logits_gender` are **raw logits**, not probabilities — the upstream PyTorch
   `forward()` softmaxes them, the ONNX export does not.
2. The gender index order is **`[child, female, male]`**.