Why accuracy alone is misleading
Benchmarks for voice deepfake detection usually report one number: error rate. But two models with similar accuracy can differ by orders of magnitude in size — and size determines latency, cost per stream, power draw and whether the model can run on a device at all. A model that is marginally more accurate but a hundred times larger is often the worse engineering choice.
The formula
The efficiency index combines both dimensions:
Efficiency index = (100 − Synthesis Error %) × (100 / Model Size in M)The first term is synthesis-detection accuracy. The second scales it by model size in millions of parameters, so a model that is ten times smaller at the same accuracy scores ten times higher.
How current models compare
| Model | Efficiency index | Synthesis error % | Model size (M) |
|---|---|---|---|
| Vosco Nano | 1,400 | 4.6 | 7 |
| Vosco Precise | 194 | 3.1 | 50 |
| Modulate | 30 | 1.6 | 316 |
| Hiya | 10 | 2.3 | 1,000 |
| Resemble | 3 | 2.1 | 3,000 |
The larger commercial models report lower synthesis error, and that matters for some use cases. The index makes the trade-off visible: Vosco models deliver competitive accuracy at a fraction of the size, which translates into lower inference latency, lower cost and lower power consumption. Full results, including replay error and speaker accuracy, are on the research page.
What the index does not tell you
No single number captures everything. The index focuses on synthesis detection; it does not include replay detection, speaker verification, robustness to noise or far-field audio, or performance on voice engines a model has never seen. We only compare against other vendors where datasets, operating points and methodology are genuinely comparable, and we publish methodology notes alongside results.
Why it matters for Physical AI
A robot or smart device has a fixed processor and a power budget. A 7M-parameter model that runs in under 10 ms on an NPU can live on the chip already in the bill of materials; a 1,000M-parameter model cannot. That is the practical reason efficiency is the first number we report — see why voice trust must run on-device.
Frequently asked questions
What is the efficiency index?
The efficiency index measures synthesis-detection accuracy per million parameters: (100 − synthesis error %) × (100 / model size in millions of parameters).
Why does model size matter for voice deepfake detection?
Model size drives latency, cost per stream and power consumption, and determines whether a model can run on a device at all.
Which Vosco model is the most efficient?
Vosco Nano, with an efficiency index of 1,400 at 7M parameters. Vosco Precise scores 194 at 50M parameters with lower error.
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.