Transcription¶
Milestone 3 — this is the design, not what runs today
Nothing on this page is implemented. transcribe and correct_transcript read needs a
plugin in this version, which is the truth rather than a stub that appears to work.
On the user's own machine, with no third-party service. Four stages, each replaceable by a plugin that declares the same activity.
1. Decode and detect speech¶
Find where someone is talking, so silence is not transcribed. On a two-hour interview with long pauses this is most of the saving.
2. Recognise¶
Whisper, through faster-whisper — the same model reimplemented on CTranslate2, which is MIT, installs from PyPI, runs on a CPU, and whose models are not licence-gated. A fresh install can therefore transcribe without an account, a key, or anything leaving the machine.
The model size is the user's choice against their hardware. Podlibre suggests one based on what the machine can do and says plainly what the trade is: smaller is faster and wronger.
3. Attribute¶
Cluster the voices into speakers, which the user then names once per episode. Diarisation is the stage that turns a wall of text into something a reader can follow, and it is also the stage with the most variable quality — crosstalk is genuinely hard.
4. Correct¶
Ordered pattern rules from the user's dictionary. No model is involved. A podcaster who says Castopod forty times a season should fix the spelling once, for good, and watch it apply everywhere.
A rule is a pattern, a replacement, and a scope. Rules live in the show file, so they travel with
the show, and a rule can be global or belong to one show. Suggestions are shown and never applied
silently: the Edit screen highlights each one, Tab moves between them, and Accept all exists
for when you trust your own dictionary.
What the user sees¶
Transcription is a task like any other: progress, an estimate, and a Stop button. It runs on a worker thread, like everything slow, so the window keeps drawing while it works.
A thread is not a process, and the difference is worth stating here rather than discovering it: a plugin shares Podlibre's process, so nothing outside the transcriber can kill a model that wedges. Stop reaches the loop and nowhere else, which makes how often that loop looks at it the whole of what Stop means here — see the architecture.