Sunday roundup: four posts on audio debugging and work culture
Four posts from a short publishing week. Tue through Fri covered audio debugging, speaker diarization, remote work culture, and a tool I reach for more than any SDK. Here is what I wrote and what stuck with me.
Tuesday: TIL on afinfo for audio inspection
I wrote about afinfo, the macOS command that reads audio file headers and tells you everything about the format before you send it to an API. Sample rate, channels, bit depth, duration, codec. In one command you catch format mismatches that would otherwise produce garbage transcripts and zero error messages. This is the kind of thing I reached for constantly at Deepgram when someone in the Discord said the model was broken. It almost never was. The audio was wrong.
Wednesday: why speaker diarization is hard
A deeper look at the hardest problem in voice AI that nobody warns you about. Transcription is a bounded language problem. Diarization is an unbounded clustering problem where the model has to figure out how many speakers exist, track them through silences, and handle overlapping speech without splitting one person into three or merging two into one. The failures are silent too, which makes them dangerous for downstream analytics.
Thursday: Slack is your remote team’s social media
The post that surprised me most this week. I wrote about how remote teams build culture through Slack channels that look like noise from the outside but are actually the only social layer remote workers have. A lot of people resonated with this one. More than I expected. If you manage a remote team and you are thinking about tightening Slack policies, read that post first.
Friday: ffmpeg taught me more than the docs did
ffmpeg is not a voice AI tool, but it is the most useful one I have. The post covered how ffprobe reveals what audio files actually contain versus what they claim to be, and the one-liner I use to resample anything to 16 kHz mono before sending it to an STT API. Format mismatch is the most common integration failure and ffmpeg catches it in milliseconds.
What I am still thinking about
The Slack post landed differently from the technical ones. It pulled in a different audience and the reactions were more personal. People do not talk enough about what remote work isolation actually feels like year after year. I will probably write more about that.
The Monday and Saturday gaps this week were the result of travel and a shifting schedule. Both days had posts in draft that did not make the edit window. They will land next week.
Frequently asked questions
What is afinfo and why does it matter for STT debugging?
afinfo is a built-in macOS command that prints audio file properties, including sample rate, channel count, bit depth, and duration. It catches format mismatches before an API call is wasted. Full post at https://lukeocodes.dev/til-afinfo-audio-inspection.
Why is speaker diarization harder than transcription?
Transcription maps audio to words, a bounded language problem. Diarization is a clustering problem where the model does not know how many speakers exist or what they sound like ahead of time. Full explanation at https://lukeocodes.dev/why-speaker-diarization-hardest-problem-voice-ai.
Does the Slack culture post apply to any remote team?
It applies to any team where Slack or similar chat tools are the primary communication layer. The dynamics around informal channels, humor as social glue, and the isolation of remote work are not specific to any one company. Full post at https://lukeocodes.dev/slack-is-your-remote-teams-social-media.
How does ffmpeg help with STT API integration?
ffprobe (part of ffmpeg) reveals the true format of audio files, and ffmpeg itself can resample, re-encode, and trim silence before sending audio to an API. The one-liner that converts any file to 16 kHz mono is in the post at https://lukeocodes.dev/ffmpeg-voice-ai-debugger.