Rating methodology
Version 2026.06 · released
| Axis | Weight | What it measures |
|---|---|---|
| Accuracy (WER) | 20% | Word error rate on real-world audio |
| Privacy / on-prem | 20% | Can audio stay on your own infrastructure? |
| Diarization | 10% | Speaker separation quality |
| RAG / agent-readiness | 20% | Vector-DB / MCP / agent integration |
| Integrations | 10% | Breadth of ecosystem connectors |
| Pricing / value | 10% | Cost relative to capability |
| Ease of use | 10% | Setup + day-to-day usability |
How we score
Every tool in the catalog is rated 0–10 on seven axes. The overall score is a fixed weighted average — the weights are published (see the table above) and do not change per tool.
Category leaderboards (for example “best on-prem transcription”) rank by the single axis that defines the category, so a tool can only top a category where it genuinely leads on the underlying axis — not by editorial fiat.
We do not claim to be “independent” or “unbiased”. We publish the rubric and the data so you can judge the scores yourself. Tool records are dated and sourced from vendors, with the check date shown on every profile.