Why speaker models matter
Every voice identity decision starts with a speaker embedding — the voice vector that represents how someone sounds. If that representation is weak, every downstream signal inherits the weakness: verification thresholds become loose, identification confuses similar voices, and spoofing detectors have less to work with.
ReDimNet and ReDimNet2
ReDimNet (2024) and ReDimNet2 (2026) are speaker-recognition architectures co-authored by members of the Vosco team and released as open source under the MIT licence, with code and pretrained weights. They are designed to lead public evaluations while staying among the smallest models in parameters and FLOPs.
The original ReDimNet code is available on GitHub.
Accuracy versus cost: the Pareto front
The useful way to compare speaker models is not a single accuracy number but a curve: error rate against inference cost. A model is on the Pareto front when no other model is both more accurate and cheaper. Across seven scale points (B0–B6), ReDimNet2 defines that front, ahead of families such as ECAPA, CAM++, NeXt-TDNN, ResNet and large self-supervised models like WavLM and W2V-BERT 2.0.
That matters in practice because you can pick the scale point that fits your hardware — a large variant on a GPU server, a small one on a device NPU — and stay on the best available trade-off.
From research to product
Vosco runs one training pipeline, one architecture and one team. The same research ships twice: as Vosco Precise behind the cloud API (99.3% speaker accuracy) and as Vosco Nano on the device (98.4% at 7M parameters). See the full benchmark table on the research page and the reasoning behind size-aware comparison in the efficiency index, explained.
What comes next
Our research programme continues on synthetic-voice detection across seen and unseen TTS engines, replay and liveness robustness, latency on representative hardware, and noise and far-field conditions for devices that hear speech across a room.
Frequently asked questions
What is ReDimNet?
ReDimNet is an open-source speaker-recognition architecture, published in 2024 and followed by ReDimNet2 in 2026, co-authored by members of the Vosco team and released under the MIT licence.
What does Pareto front mean for speaker recognition models?
A model is on the Pareto front when no other model is both more accurate and cheaper to run. ReDimNet2 defines that front across seven scale points.
Is ReDimNet used in Vosco products?
Vosco’s models come from the same research line and training pipeline, delivered as Vosco Precise in the cloud and Vosco Nano on the device.
Verify the voice before your AI acts.
Start on the Vosco API with 1,000 minutes free every month, or talk to us about on-device deployment.