Request for non-quantized pre-trained models

[yf-fan-org]

Is it possible to provide a non-quantized model, as I have found that quantized models sometimes don't work in real tests, and I'd like to test how accurate non-quantized is

[yf-fan-org]

My native language is not English, and sometimes it's hard to detect when the pronunciation is not quite right, and I wondered if I could improve it by adding some additional audio data. Thank you very much for your answer!
@kahrendt

[Nold360]

My native language is not English, and sometimes it's hard to detect when the pronunciation is not quite right, and I wondered if I could improve it by adding some additional audio data. Thank you very much for your answer! @kahrendt

I wonder if fine-tuning the models would be possible.. 🤔 maybe that would help?
#38

[kahrendt]

In my benchmarking tests, the quantization doesn't reduce the model accuracy in a consistent way. Sometimes it performs slightly better than the floating point model and sometimes slightly worse. When I mean slightly different, I mean the probability cutoff to only have a 5% false rejection rate differs from by 0.01.

There isn't an easy way to run the non-quantized models; they won't work on an ESP32. They really only work using the test functions in the mWW code itself.