Performance degradation on MS Teams with camera + background effects (Whisper Large)

[devin-ai-integration]

Description

A user reported significant performance degradation when using Hyprnote on a MacBook Pro 16" M1 (16GB RAM) while on MS Teams with camera on and background effects enabled.

Environment

• Device: MacBook Pro 16-inch M1, 16GB RAM
• Video call app: MS Teams (camera on + background effects)
• Models: Whisper Large v3 + Llama 3.2 (running locally)
• Previous version: 0.0.84 (was running much smoother with the same setup)

Details

• The user was previously on v0.0.84 (beta) where everything ran smoothly with the same hardware and setup.
• After upgrading, everything looks slow — likely due to the switch from whisper.cpp with batch processing to argmax + parakeet (real-time processing).
• With Parakeet 3, there are fewer issues, but the camera still looks choppy.
• The user suggests it would be great to:
  • Implement leaner transcript models
  • Support processing transcription after the meeting is over (post-processing mode)

Steps to Reproduce

1. Use a MacBook Pro M1 16GB
2. Join an MS Teams meeting with camera on and background effects enabled
3. Use Whisper Large transcript models locally
4. Observe performance degradation / choppy camera

Notes

• Parakeet 3 helps reduce the issue but does not fully resolve the camera choppiness.
• This was reported in Discord #dev channel on Feb 9-10, 2026.

[devin-ai-integration]

Additional Context

From the continued Discord conversation:

• Kpiquax5 explains: Teams' virtual background is a resource hog — it runs a segmentation model in real time, frame by frame, to separate you from the background. That on top of Whisper transcribing in real time is just too much for the Neural Engine/GPU.

• John suggests: Have an option in Hyprnote settings to keep the recording and then batch process afterwards for people in this situation.

Proposed Solution

Add a batch processing mode option in Hyprnote settings that allows users to:

1. Record audio during the meeting without real-time transcription
2. Process the transcription after the meeting ends

This would help users whose GPU/Neural Engine is already under heavy load from other apps (e.g., MS Teams virtual backgrounds).