The developers of the BOLTUN communication platform are testing noise suppression solutions to improve audio quality while minimizing CPU usage. Current experiments focus on two main approaches: VoxLock for voice isolation and a lightweight threshold-based processor for handling pauses.
VoxLock uses a six-second voice sample to isolate a specific speaker. Initial tests on an Intel Core i5-11400H processor showed varying performance, with processing times for a 12-second audio clip reaching 25.92 seconds in early prototypes. Further optimizations reduced latency, allowing 3750 blocks to be processed within a 250 ms window during a one-minute stress test.
The team also evaluated a server-side processing model to offload computation from the user's PC. While this setup handled up to five simultaneous streams efficiently, it failed with ten continuous streams. Consequently, the team shifted focus toward a lightweight threshold-based DSP that manages signal levels during silence without heavy neural network overhead.
The new threshold-based tool features a 40 ms opening trigger, a 240 ms hold time, and a 100 ms pre-roll buffer to preserve the start of words. Unlike VoxLock, this method does not isolate voices from background noise but rather gates the transmission. Future development will include real-world testing during gameplay to balance noise reduction with system performance.
