Actual accuracy / false-positive results for the hey_echo example model?

[karthikeyareddy-create]

Thanks for putting this together — the granular step pipeline and custom_negative_phrases support are exactly what I needed.
I noticed the README documents target_false_positives_per_hour: 0.2 as the default config target, but doesn't show what was actually achieved when you trained the hey_echo example model. Could you share:

• Final accuracy/recall on your validation set
• Measured false positives/hour (and how/where it was measured — e.g. against the ACAV100M negative set, real-world background audio, etc.)
• Did you use the generic negative data as background clips for augmentation of the data?
• Roughly how many n_samples / training steps it took to hit that target FP rate

Also — the README mentions noise augmentation in step 9 but doesn't specify the SNR/dB range used when mixing background noise into clips. Is this exposed as a config option, or hardcoded? If hardcoded, what range (in dB SNR) is applied?

Would help a lot in setting expectations before a full training run. Thanks!