I'm currently evaluating real-time noise suppression solutions for a series of high-attendance webinars (500+ concurrent attendees) we are migrating to a live, interactive format. The primary platform is a custom WebRTC implementation, though we also use standard conferencing software as a backup. Audio clarity for the presenter is the non-negotiable metric, as background noise (keyboard, HVAC, domestic environments) has been a significant point of feedback in our post-event surveys.
My requirements are stringent: the solution must provide consistent, high-fidelity suppression without introducing artifacts or latency that would disrupt the natural flow of Q&A. It must also be operationally stable under sustained load for 60-90 minute sessions. I have narrowed the field to several contenders, including Krisp, RTX Voice, and a custom RNNoise implementation. Krisp is frequently recommended, but I lack concrete data on its performance at this specific scale and in a controlled, professional production environment.
I am seeking detailed, technical reviews from members who have deployed Krisp in analogous large-scale, real-time scenarios. Benchmarks or subjective comparisons on the following parameters are particularly valuable:
* **Latency Introduced:** Measured end-to-end delay added by the processing pipeline. Even 40-60ms can become noticeable.
* **CPU Utilization:** Profile on a modern CPU (e.g., AMD Ryzen 9 7950X or Intel i9-13900K) during a simulated 2-hour session. We need to ensure no thermal throttling or contention with other production software (OBS, video encoders).
* **Artifact Analysis:** Specific noise types where the algorithm struggles. Does it overly attenuate plosive sounds ('p', 'b') or cause "underwater" effects on certain voice timbres?
* **Failover & Stability:** Behavior during network packet loss or system resource exhaustion. Does the audio stream cut out, or does it gracefully degrade?
* **Configuration Nuances:** Optimal settings within the Krisp application for a dedicated broadcast machine. For example:
```
// Are there CLI flags or config file tweaks beyond the GUI?
--mode broadcast
--aggressiveness 0.85
--sample-rate 48000
```
* **Comparative Efficacy:** A/B test results against other solutions, especially for non-stationary noises (clicky mechanical keyboards, distant traffic, appliance hum).
Our current baseline is a hardware gate/compressor combo, which is effective but inflexible for our remote presenters. The promise of a software-based solution like Krisp is compelling from a scalability and cost perspective, but the risk of audio degradation at peak load is a serious concern. Please share any performance telemetry, oscilloscope/spectrogram screenshots, or structured observations you may have.
Data over dogma