Overview

This month has been a blend of polishing Nucleate’s user experience and tackling technical hurdles, particularly around cross-platform compatibility. I focused on enhancing features, especially around transcription and diarization, while also addressing some persistent issues on the Mac side. The month was marked by a shift in testing approaches and a deeper integration of various audio processing technologies.

Highlights & Progress

Tools & Workflow

• Polished the app’s stylization and resolved several Mac-related issues, which has been a long-standing pain point.
• Created a single point of contact theme class for easier customization, streamlining the styling process.
• Updated the website and documentation materials for Nucleate’s features.

Systems & Features

• Added new features like Ollama detection, an Ollama overdrive toggle, and gated diarization behind pro features.
• Addressed additional bugs specifically affecting the Mac version of the app.
• Integrated OpenAI’s Whisper and Faster Whisper for transcription, as well as resolved diarization subprocess woes in the app.
• Implemented M4A file conversion into .wav and normalized backend transcription processing.

Experiments

• Successfully tested diarization via pyannote audio on CPU and on GPU.
• Upgraded to the GPU package for torch and identified CUDA version 11.1 for GPU compatibility, although I faced some limitations with Torch options on Mac.

Roadblocks

• Encountered inconsistencies in app performance between Windows and Mac, which has been frustrating and requires further investigation.
• Faced dependency issues on Mac, leading to a workaround using a virtual environment and CPU-only Torch builds on Intel Mac hardware.
• OpenAI Whisper’s incompatibility with Intel Macs led to the decision to make Faster Whisper available to all users, which feels like a necessary compromise.

Reflections

Overall, I feel like I made solid progress this month, especially in terms of feature development and improving the user interface. However, the ongoing issues with Mac compatibility have been a significant roadblock. I was surprised by how much I had to pivot my approach to testing and development, particularly with the CPU-based strategies. The decision to gate certain features behind a pro tier is still up in the air, as I’m weighing the benefits of making hardware-accelerated transcription more widely available.

Next Steps

• Continue addressing the ongoing bugs on the Mac side and investigate the VRAM spike during diarization.
• Finalize decisions around the accessibility of hardware-accelerated transcription and the gating of features.
• Reach out to potential testers on Reddit to form a small beta group for user feedback.
• Enhance the reliability of diarized insights and explore when to create marketing animations for the app.

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