A software engineer transcribed ten hours of archived podcast interviews overnight on his laptop using free open-source AI models, a task that would have been impractical in 2016.

What happened

The author used WhisperX (which wraps OpenAI's Whisper large-v3 model) and pyannote.audio (for speaker diarization — splitting audio by speaker) running locally on his Apple Silicon laptop to transcribe and label a ten-episode podcast archive. The work took roughly twice real-time per episode (about 14 hours of compute total) and produced searchable, timestamped transcripts with speaker labels for each conversation.

Why it matters

In 2016, decent speech-to-text was a paid cloud service and speaker identification was a research-level challenge; transcribing ten hours of interviews yourself was not realistic. The same end-to-end task now runs locally, without API bills or leaving the machine, requiring only a free Hugging Face account and an evening of time. This marks a concrete shift in what one person can do with open-source AI tools on consumer hardware.

What to watch

Both WhisperX and pyannote.audio run on CPU with int8 quantization and require no GPU, though performance on Apple Silicon is slow (roughly twice real-time). The diarization models are gated on Hugging Face and require accepting license terms before download. The final transcripts are now published on episode pages with clickable timestamps linking to the audio.