Quick Guide
How to Improve Accuracy
Practical tips and best practices to help you get the best possible voice recognition performance out of BenASR.
-
1
Get a good microphone
Microphone quality matters a lot, especially if you are in a noisy room or a room with echo. You don't need an expensive microphone. We recommend all USB microphones; Sennheiser PC 7 is a well-tested, all-around budget option.
-
2
Record more samples
The more samples you record, the better the model will get. If you use the default word list with 110 words, the model gets close to saturation around 20K samples. Before that point, keep recording more samples, especially for words the model has a hard time detecting.
-
3
Do not forget sample cleanup
Even a few bad samples, such as noisy clips, wrong spoken words, or clipped audio, can affect accuracy. Listen to your recorded samples and remove unsuitable ones.
-
4
Use real-world samples instead of random words
Real usage gives the model better examples of your workflow. BenASR stays local for daily recognition, and we do not collect your private data or send it off your machine.
-
5
Make sure your pronunciation is correct
Some words may not match the American lexicon used by the engine. In Train Studio hover on each word to check its phonetics, then keep pronunciation consistent so the model does not learn conflicting sounds for the same phonetic.