DRE currently offers 3 main speech recognition modes to choose from in `Sound -> Recognition -> Speech Input Engine`:

1. Classic: Local Windows Speech Recognition
   1. This uses existing voice command grammar to match up with your voice inputs.
2. Neural Speech Recognition
   1. This uses neural models to process and transcribe your voice inputs.
   2. A custom AI, Thyrring, is then used to match your voice input intents to the available commands.
   3. This mode can be run locally using your GPU/CPU, or in the Cloud (Performance tier or higher).
3. Hybrid
   1. This is a combination of Classic and Hybrid, yielding the best results.

Here’s the modes and sub-modes order of least to most impact on your computer:

1. Classic with Microsoft Runtime
   • This uses a local, basic model with no ways of training it to your voice. This is the least precise option in terms of matching your speech to what you want.
2. Classic with Windows Runtime
   • This uses a local model with ways of training it to your voice (Windows 10 only). This is slightly more precise because it adapts to your voice through specific training and every day use. If you started using DRE on Windows 11, you should use the Microsoft option.
3. Neural with Cloud Provider
   • This uses a cloud transcription service to transcribe your voice inputs. Since it’s on cloud, your PC isn’t used for this. When processing the transcription to match it to any of the voice commands, there’s a small CPU impact, but mostly negligible
4. Neural with Local Provider
   • This uses a local download model to transcribe your voice inputs. The `Sound -> Recognition -> Preferred Neural Model` option decide the impact
     • Small model: Yields good results. Good for older PCs.
     • Medium model: The best balanced model, yielding great results, but has a larger impact on your system.
     • Large model: The perfect model, but has the largest impact on your computer.
     • Usually Medium is the preferred option, but dropping to Small can work great with the Hybrid mode
   • Playing around with the order of `Sound -> Recognition -> Preferred Neural Hardware Priority` may help in your situation.
5. Hybrid
   • Uses the Classic recognizer first to detect direct grammar matching before sending it off to transcription
   • This mode can be efficient on your system, as long as you keep the Neural provider to Cloud, or a Small/Medium local model

So what should I choose? Well it depends on your system. Got a beefy GPU? You can likely handle a Local Medium Neural model. Run in VR though? You’re better off with a Small model or subscribing to Performance to unlock the Cloud option.