Transcribe
You point it at a recording. It asks Gemini for a verbatim transcript, then writes a private folder: manifest, transcript (JSON, text, and SRT), speaker labels, timestamps, audio observations, summary, topics, decisions, actions, questions, speaker-name notes, and a translation.
The transcript stays in the original language. Translation defaults to English. Speaker labels such as SPEAKER_01 mark voices inside that one file. They are not identified people.
How a run moves
flowchart TB
audio[One audio file]
snap[Private temp copy of those bytes]
up[Upload to Google Files API]
tx[Transcribe and note how it sounded]
save[Save the transcript]
enrich[Summarize, topics, translation]
bundle[Manifest plus the files]
del[Try to delete the uploaded file]
audio --> snap --> up --> tx --> save --> enrich --> bundle
up --> del
tx --> del
If the summary step fails, the transcript already on disk is kept. Resume continues from the manifest and does not upload the audio again.
sequenceDiagram
participant CLI as gemini-transcribe
participant Google as Gemini
participant Disk as Output folder
CLI->>Disk: snapshot the input bytes
CLI->>Google: upload, then transcribe
Google-->>CLI: transcript and observations
CLI->>Disk: write transcript
CLI->>Google: analyze and translate the saved transcript
Google-->>CLI: summary, topics, translation
CLI->>Disk: write the rest of the bundle
CLI->>Google: delete the uploaded file
Supported inputs include WAV, MP3, AIFF, AAC, OGG, FLAC, and M4A, under 2 GiB. The whole file goes in one pass. Long recordings can still exceed the model's limits. SRT times are the model's estimates.
The default model is gemini-3.8-flash. An optional --pipeline long-decode uses gemini-4-argon and can continue generation across requests. Both pipelines write the same kind of bundle.
Standard requests set interaction storage off (store=False). That is an opt-out of that storage, and it leaves Google's other data terms as they are.
What the folder is for
WikiContext can later ingest a bundle as evidence. This tool does not publish a wiki page, and a finished extraction is not a reviewed publication. Timestamps can overlap. Uncertain and inaudible spots are supposed to be marked. A summary can be wrong even when the files are well formed.
Audio under audio/ and output under transcripts/ stay out of Git.