transcribe_

Terminal audio workbench · v0.4.0

Long recordings, split without losing the thread.

A Textual app that preflights long audio, keeps chunk timestamps aligned, returns speaker labels for review, and exports text or SRT from the terminal.

Run the preflight Inspect the source ↗
transcribe · release-interview.m4a
TRANSCRIBE_● preflight complete
recording
duration25m 00s3 bounded chunks
pipeline
split
3 / 3
upload
3 / 3
transcribe
2 / 3

Free key · work stays sequential · remaining request pressure stays visible.

10 minmaximum chunk boundary in current source
3 phasesseparate split, upload, and transcribe progress
5 maxconcurrent transcription workers for paid-key mode
3 outputsplain text, timestamps, or SRT subtitles

00 / product demo

From local guard to reviewed export in nine seconds.

The recording exercises the same deterministic browser proof below: a local free-key guard, a paid-key concurrency preview, a speaker rename, and SRT output. Action callouts keep every transition explicit.

DETERMINISTIC BROWSER PROOF · SANITIZED TEXT · NO AUDIO UPLOAD · NO MODEL REQUEST

01 / deterministic preflight

See the stop condition before the model call.

Change duration, entered session usage, or key tier. This sandbox mirrors the repository’s chunk and request-pressure rules without uploading audio or calling Gemini.

Key tier

BROWSER-ONLY PROOF · NO AUDIO UPLOAD · NO API KEY · NO MODEL REQUEST

preflight result

Review quota
3chunks
16local slots left
1workers
01 split
ready
02 upload
ready
03 transcribe
ready

02 / human review

Speaker labels stay editable before export.

Rename the sanitized demo speakers and switch output formats. The terminal app also exposes audio preview for speaker identification; this browser proof keeps the interaction silent and deterministic.

Detected speakers
Export preview 

        

      

    


    

      

        
03 / system path
A visible pipeline, not a single spinner.

        
The application separates preparation, provider work, and human review. Missing chunks are surfaced as errors instead of being embedded into a transcript as if they succeeded.

      

      
      
FAILURE PATH · provider quota exhaustion stops remaining free-key work; upload and transcription failures keep their part and phase visible.

    


    

      

        
04 / source-rendered interface
The terminal app is the product.

        
This is the current Textual result screen rendered from public source with a sanitized 24-second transcript. The text demonstrates interface behavior; it is not presented as a completed Gemini transcription.

      

      
        Open full-size interface ↗
        
Transcribe Textual result screen with output choices, sanitized speaker transcript, save path, and completion summary

      
      

        
Output and transcript · scroll horizontally
Full-size crop of the Textual output formats and sanitized transcript

        
Save and completion state · scroll horizontally
Full-size crop of the Textual save controls and completion summary

      

      
SANITIZED DEMO DATA · CURRENT PUBLIC TEXTUAL UI · NO CLAIM OF MODEL ACCURACY OR CUSTOMER USE

      

        
Local prerequisites
Python 3.10+, uv, ffmpeg, and ffprobe are required to run the terminal app.

        
Provider boundary
The application uses a user-supplied Gemini API key. The local guard cannot read project-level provider usage; limits and output quality are not guaranteed here.

        
Demo boundary
This page tests deterministic planning and formatting only. It does not accept files or send requests.