Minimum Required Technical Specifications - Myna Documentation

[dmxplx]

Hi folks,

One thing I haven’t seen mentioned in the Discourse article yet is the minimum technical specifications required for the project (apart from here: https://github.com/canonical/myna/tree/main/specs).

I have a recent machine with a good audio setup, and I believe I might not have issues in this area, even though others might. That’s why I’m opening this discussion. I’m starting from the assumption that every installation will be eligible, provided there is enough storage to install all the required dependencies, which took quite a bit of space on my end.

Are there any minimum requirements regarding CPU, RAM, GPU, storage, or other hardware specifications (USB Mic, etc...)? It would be useful to know what the baseline requirements are for the project (unless I missed them).

Thanks

[quine00]

Hi Dimitri,

We have plans to run performance benchmarks against a wide range of hardware in the Canonical labs to attach grounded recommendations. For now, it has only been ad-hoc testing on a smattering of dev machines.

We'd also like to introduce a feature that will benchmark your machine at setup time and recommend to you whether you'll be able to run these local models effectively. For example, transcribing a reference clip and measuring the so called "real-time factor". If that is << 1, we suggest you won't have fun using Myna, for instance.

Thanks for your thoughts and questions!

[quine00]

We also have some beginnings of mic evaluation. E.g., if your mic is very noisey, Myna can detect that in certain cases and warn you about it. Noisey inputs can push the models into pathologically bad transcriptions, so it is important to have a properly configured mic.

[dmxplx]

Thanks for the insights behind the scenes @quine00. As mentioned above, I recently debugged this, if that’s helpful to others on the topic, particularly for improving the quality of audio acquisition for Myna. https://discourse.ubuntu.com/t/how-to-usb-mic-at-192-khz-on-ubuntu-without-killing-bluetooth-audio/88376

We also have some beginnings of mic evaluation. E.g., if your mic is very noisey, Myna can detect that in certain cases and warn you about it. Noisey inputs can push the models into pathologically bad transcriptions, so it is important to have a properly configured mic.

[quine00]

Thanks, useful for getting that class of mic working, though it won't improve Myna's accuracy.

The models we've evaluated so far were all trained on 16 kHz audio, and their input stage (a log-mel feature extractor) only looks at 0-8 kHz. That's why Myna captures at 16 kHz: anything a 192 kHz source adds above 8 kHz gets removed during resampling.

Forcing the PW graph to 192 kHz can also mean resampling every other stream, which costs CPU and power, so I wouldn't bother unless you need it for something else.

What actually matters is SNR (a closer or directional mic), correct input gain (no clipping, not too quiet), and not stacking noise suppression or auto-gain on top of each other. Sample rates above 16 kHz and bit depths above 16 bits make no difference.

[dmxplx]

Super insightful, thanks! I’m surprised to hear that Myna captures at 16 kHz. Have a great day!