VoxProfile

Predicts a speaker's gender and age decade from a short audio clip. A pretrained speaker-embedding encoder (TitaNet-L) feeds two lightweight classification heads, so both attributes come out of a single forward pass.

• Gender: binary classification (Male / Female)
• Age: 6-way decade classification (10s 20s 30s 40s 50s 60s+)

Code: furence-ai/VoxProfile

Architecture

1. Raw waveform goes through a pretrained TitaNet-L speaker-recognition encoder, producing a 192-dim speaker embedding (encoder frozen by default; top layers can be unfrozen, or adapted with LoRA, instead).
2. The embedding feeds two independent MLP heads, one producing a single gender logit and the other 6 age-decade logits.
3. Both losses are combined and trained jointly: loss = BCEWithLogitsLoss(gender) + λ × CrossEntropyLoss(age)

Performance

Evaluated on a 624-sample held-out test set.

Metric	Value
Gender accuracy	98.9%
12-class joint accuracy (gender AND age decade both correct)	68.8%

Age decade	n	MAE	Decade acc
10s	97	0.825	60.8%
20s	113	0.212	79.6%
30s	108	0.565	47.2%
40s	97	0.567	45.4%
50s	73	0.247	82.2%
60s+	136	0.257	94.9%

Gender is nearly solved (well under 1% error). Age confusion concentrates in adjacent decades (30s↔40s, 20s↔30s), while the extremes — 10s and 60s+ — separate cleanly from the rest.

Training data

Utterances with gender and age labels were drawn from several Korean-language AIHub corpora (free conversation, meeting recordings, speaker-recognition audio), cleaned, and split by speaker into train/val/test.

Usage

Requires the model code from the GitHub repo (voxprofile/model.py, voxprofile/config.py) — this repo hosts only the trained weights.

import torch
from voxprofile.config import TrainConfig
from voxprofile.model import AgeGenderModel
from huggingface_hub import hf_hub_download

cfg = TrainConfig()
ckpt_path = hf_hub_download(repo_id="furence-ai/VoxProfile", filename="model.pt")

head_cfg = cfg.head_config()
model = AgeGenderModel(
    encoder_name=cfg.encoder_name,
    encoder_out_dim=cfg.encoder_out_dim,
    hidden_dims=head_cfg["hidden_dims"],
    dropout=head_cfg["dropout"],
    unfreeze_encoder_layers=head_cfg.get("unfreeze_encoder_layers", 0),
    emb_dropout=cfg.emb_dropout,
)
model.load_state_dict(torch.load(ckpt_path, map_location="cpu"))
model.eval()

License

MIT