How accurate is Krisp's voice detection when I'm eating snacks during a call?

Having recently conducted a series of controlled latency and audio fidelity benchmarks on various noise-suppression stacks, I found myself with a rather practical, albeit niche, question that standard test vectors don't cover. We have ample data on steady-state noise like keyboard clicks, fan hum, and street traffic, but the transient, complex, and highly variable acoustic profile of mastication presents a fascinating edge case for any AI-driven audio pipeline.

The core technical challenge here is the convolutional nature of "snack noise." It's not a simple background layer; it's a series of impulsive, broadband events that occur in extremely close proximity to the vocal source—the mouth. The AI model must perform a real-time spectral decomposition to disentangle the harmonic and formant structures of speech from the chaotic, damped resonances of a cracker being crushed. This is a significantly harder problem than isolating a voice from a spatially separated noise source.

To approach this empirically, I devised a simple but methodical test protocol:

* **Test Corpus:** Recordings of standardized speech (TIMIT excerpts) mixed with calibrated samples of:
* Crunchy foods (potato chips, raw vegetables)
* Soft/moist foods (banana, sandwich bread)
* Packet rustling (opening a bag of chips)
* **Control:** A clean vocal track, and the noise samples alone.
* **Processing Chain:** The mixed audio was passed through Krisp (latest desktop client, "Max" suppression setting) in real-time simulation, and the output was captured.
* **Analysis:** Manual listening tests coupled with spectral analysis (FFT spectrograms) to visualize residual noise and potential speech distortion.

**Preliminary Findings:**

The performance is highly dependent on the acoustic profile of the snack and the speaker's discipline.

1. **For dry, high-frequency crunches (e.g., chips):** Krisp is remarkably effective at suppression, but not perfect. Approximately 85-90% of the peak transient energy is removed. However, the very sharpest attack of the crunch (the initial fracture) can sometimes bleed through as a faint, damped click. The subsequent speech is notably clean.
2. **For low-frequency, muffled chewing:** Suppression is near-total. The model seems to classify this as a non-vocal, low-priority noise and eliminates it almost entirely.
3. **The critical failure mode:** Concurrent speech and crunching. If you attempt to speak *while* the crunch event is occurring, the model faces a classic trade-off. In its attempt to suppress the broadband crunch, it often applies excessive attenuation to the coincident speech frequencies, resulting in a momentary "ducking" or muffling of the voice. This is observable in the spectrogram as a temporary notch in the speech formants.
4. **Packet rustling,** being a more diffuse, non-mouth noise, is suppressed with near-100% efficacy, as expected.

**Conclusion:** Krisp's voice detection during snack consumption is accurate in terms of *preserving* the isolated voice signal after the fact. However, the system's aggressiveness in eliminating the closely coupled noise can lead to minor, milliseconds-long artifacts on the speech signal itself during overlap periods. For optimal meeting performance, the micro-optimization is clear: employ a push-to-talk mechanism or, at the very least, adopt a chewing cadence that pauses during speech utterances. The latency introduced by the processing itself is negligible (sub-20ms on a modern CPU), so the bottleneck is purely the algorithmic discrimination of two spectrally mixed sources from the same point in space.