I spent a good chunk of time testing packaging across different platforms and hardware, focusing mainly on Whisper and Faster Whisper. One thing that stood out is that Apple Silicon Macs handle Faster Whisper pretty well since it doesn’t depend on Torch, which is a relief. Whisper itself is still giving me trouble after the recent refactor—I found some bugs in the MPS and CPU backends during processing, and there are still unresolved issues that need more testing and fixing. On Windows, there might be a problem with Whisper related to FFmpeg, and on Mac, subprocess launches tied to FFmpeg have been failing, which complicates things.

I dug into the compatibility problems and discovered a conflict between the Torch codec and FFmpeg/ffprobe that was causing major issues with Whisper. Switching back to Torch 2.8 helped stabilize it, so that’s the route I’m sticking with for now. For Faster Whisper, I fixed timeout issues by changing how logging works—rather than printing lines directly, I switched to logging debug messages, which also helped avoid those annoying flashing black boxes by hiding all subprocess calls.

Packaging-wise, I tweaked the spec file to improve the process and managed to package the Mac app as a .dmg. On Windows, I set up an installer with a progress indicator, and the CUDA backend is working and packaged for all known backends. I still need to finish packaging Windows without CUDA and wrap up the Mac build for Apple Silicon before the beta release. The app is still incompatible with non-CUDA GPU setups on Windows, which is a problem I haven’t figured out yet.

I moved the file rebuilding option into the primary menu to make it more accessible for free users. The app size remains pretty large, especially on CPU-only setups, so I’m planning to focus on optimizing that for version 2.0. I’m also curious about how the lightweight Ollama model manages its efficiency and want to investigate that further.

On the licensing front, I decided to distribute unique, hashed licenses during sales to prevent modification and piracy. I updated the GitHub and Gumroad pages with the relevant license info to reflect this change. My brother agreed to be a pre-beta user to give me UX feedback, which should help catch issues before the wider beta.

There are still open questions I need to tackle: fully fixing Whisper’s remaining issues after the refactor, figuring out the best long-term logging approach for Faster Whisper, resolving the incompatibility with non-CUDA GPUs on Windows, and finalizing the packaging steps for Windows without CUDA and Apple Silicon Macs. I’m aiming to start beta testing around mid-March, so there’s a fair bit to wrap up before then.

Two-Line Takeaway

Key step forward

The most important forward progress step is stabilizing Faster Whisper on Apple Silicon Macs and fixing timeout issues by improving logging, as well as packaging the Mac app and setting up a Windows installer with CUDA backend support.

Significant barrier or risk

The most important barrier or risk is the unresolved compatibility problems with Whisper, including bugs in the MPS and CPU backends, conflicts between Torch and FFmpeg, subprocess launch failures on Mac, and the app’s incompatibility with non-CUDA GPUs on Windows.

Keynotes

Decisions & Observations

• Decided to stick with Torch 2.8 to stabilize Whisper due to conflicts with FFmpeg/ffprobe and Torch codec.
• Decided to distribute unique, hashed licenses during sales to prevent modification and piracy; updated GitHub and Gumroad license info accordingly.

Issues & Next Steps

• Needs to resolve remaining bugs in Whisper’s MPS and CPU backends after recent refactor.
• Needs to fix subprocess launch failures tied to FFmpeg on Mac and potential FFmpeg-related issues on Windows.
• Needs to resolve incompatibility of the app with non-CUDA GPU setups on Windows.
• Needs to finalize packaging for Windows without CUDA and complete Mac Apple Silicon build before beta release.
• Will explore optimizing app size, especially for CPU-only setups, targeting version 2.0.
• Will investigate how the lightweight Ollama model achieves efficiency for potential improvements.
• Will gather UX feedback from pre-beta user (brother) to catch issues before wider beta.
• Aiming to start beta testing around mid-March, with remaining packaging and bug fixes to complete beforehand.

Uncategorized Notes

• Fixed Faster Whisper timeout issues by switching from direct print lines to debug logging, which also hides subprocess calls.
• Moved file rebuilding option into the primary menu for better accessibility to free users.

Generated with Nucleate by Watchlight Studio — distilled from full transcription.