ReDimNet2-b6 (vb2+vox2_v0, large-margin) — Sarvam internal mirror
Private mirror of PalabraAI/redimnet2
at commit 2a8d15f65b1dfb5d73fede2f11ee42bcccca3035, MIT licensed
(© 2026 Palabra.ai — see LICENSE). Not a Sarvam model. No weights,
architecture or code have been modified; this is a byte-identical copy.
Why this mirror exists
dubbing-service's speaker-embedding role bakes these weights into its
image at build time. Fetching them from GitHub during every image build was
unbounded — no timeout, no retry — and is the likely cause of a build that
took 63 minutes against a 22-minute norm. A build should not be able to hang
on a third-party release asset.
It also pins provenance: the embedding space must be identical across every pod, because reconciliation gates on a fixed cosine threshold (0.628, calibrated against exactly these weights). A mirror we control cannot move under us.
Contents
b6-vb2+vox2_v0-lm.pt |
checkpoint, {model_config, state_dict} (~49 MB) |
redimnet2/ |
model code, needed to construct ReDimNet2Wrap |
LICENSE |
upstream MIT licence |
Usage
import sys, torch
from huggingface_hub import snapshot_download
d = snapshot_download("sarvam/redimnet2-b6-vb2vox2-lm")
sys.path.insert(0, d)
from redimnet2 import ReDimNet2Wrap
ckpt = torch.load(f"{d}/b6-vb2+vox2_v0-lm.pt", map_location="cpu")
model = ReDimNet2Wrap(**ckpt["model_config"]).eval()
model.load_state_dict(ckpt["state_dict"])
Output is 192-d; L2-normalise before comparing. Consumers should keep to 16 kHz mono, minimum 0.4 s, and cap long audio by sampling evenly-spaced sub-windows rather than truncating to a prefix.