I Analyzed Every Gong Hit on TBPN and Here is What I Found

Initial Planning: I took this opportunity to expand my Python skills. The terrifying thought of manually skimming through hours of videos for gong hits left me with one option: using machine learning, an area in which I had no prior experience and no clear starting point.

Discovery: Since I wanted to work with Python, TensorFlow was the logical choice. This led me to learn about YAMNet, an open-source audio classification model, which luckily already had a gong sound classifier.

First Test Run: On my first test run, the TBPN audio analysis system detected roughly 35% of the gong hits in my initial batch, which was not a surprise, as the gong classifier is very poorly rated. For example, YAMNet expected a smooth, traditional hum of a gong (sample from YAMNet library), whereas John and Jordi hit the gong as if it owed them money. In fact, you would not be reading this post right now if that weren't the case.

My favorite podcaster on my favorite live podcast (thumbnail).

Training: At that stage, “verification” meant dragging through 3-hour YouTube videos looking for sudden camera cuts or timestamps when a guest announced funding. I collected ~300 manually labeled samples and trained a classifier on top of the model. It worked very well with ~95% accuracy in testing, and caught every gong hit during testing (along with the occasional false positive from Jordi playing the boat horn sound effect).

Detection Process: When processing an episode, the system downloads the YouTube audio and preprocesses it for detection. The audio is then analyzed locally using a custom-trained model running on my computer, applied in 0.96-second overlapping windows. Any segment scoring at least 92% confidence is flagged as a potential gong hit. Since a single gong strike can trigger several overlapping detections, the system consolidates detections in close temporal proximity, retaining only the highest-confidence detection in each group. Each detection is accompanied by an automatically generated timestamp, which is subsequently verified by hand, an essential step for accurately labeling funding-related gong hits. The model is purposely tuned to overreport rather than underreport, as it's preferable to filter out false positives than miss actual events.

Measuring Loudness: Peak detection alone was ineffective because nearly every hit was “maxed out.” To address this, I implemented the EBU-R128 broadcast standards, using Loudness Units Full Scale (LUFS) for relative loudness and True Peak for spike measurement. Both are recognized as more accurate indicators of how the human ear perceives loudness. With these measurements, I calculated the Peak-to-Loudness Ratio (PLR), which became the core standard for the entire project. PLR was chosen because it captures the absolute loudness of the gong hit at impact (True Peak) while accounting for relative sound levels before and after impact (LUFS). Here is the primary reference I leveraged for this entire element of the project. For my analysis, I normalized PLR values across all gong hits to enable easier comparison.

Output: Results were output as .csv files containing timestamps, confidence scores, and video metadata. The pipeline processed entire episodes in minutes, and detection parameters could be easily adjusted as needed; further implementation details and code are available in the linked GitHub repository. Some hits lasted so long that they triggered multiple detections, despite my ~8 second detection consolidation mechanic when John used the “proper” form.

Manual Labeling: For manual labeling, I personally reviewed each hyperlinked video segment flagged by the detection system. For every detected gong hit, I watched the corresponding moment to verify if it was related to a funding announcement. If it was, I marked down the relevant funding data, and for every hit (not just the funding hits), I noted which host struck the gong. This careful review ensured each event was accurately categorized and attributed within the dataset.

Late August was pretty fun.

Final Dataset: The final dataset was built to export all results for every detection across all episodes since the new studio. I added manual data and built the comprehensive dataset that powers the analysis below. Between labeling, testing, and refining, I listened to a bare minimum of 1,163 gong hits during the development process. However, the actual number is likely at least double, as I was usually listening to 2-3 hits per page visit. On August 30th alone, I heard 556 gong hits.

Note: my data set spans from the launch of the new studio in late May to the end of August 2025. I essentially spent the month of September writing this post. If I had the time to do all the charts and analysis again, frankly, I would have included September, but at least anecdotally, it seems as though the TBPN team has cracked down on the audio volatility of the gong hits... for better or worse.