gemini-transcribe-wrapper

Zero-config wrapper for free-tier Gemini 3.5 Transcribe that breaks long audio into API-safe chunks and merges the results back into .speakers.srt / .srt / .txt.

• GitHub: https://github.com/tayaee/gemini-transcribe-wrapper
• PyPI: https://pypi.org/project/gemini-transcribe-wrapper/

Quick Start

Prerequisites

• Get an API key for free from https://aistudio.google.com/api-keys/.
• Install uv (https://docs.astral.sh/uv/getting-started/installation/).

Install the tool

Linux / macOS:

export GEMINI_API_KEY=your_key_here
uv tool install --python 3.13 gemini-transcribe-wrapper@latest
gtw -v

Windows (PowerShell):

$env:GEMINI_API_KEY = "your_key_here"
uv tool install --python 3.13 gemini-transcribe-wrapper@latest
gtw -v

Windows (Command Prompt):

set GEMINI_API_KEY=your_key_here
uv tool install --python 3.13 gemini-transcribe-wrapper@latest
gtw -v

Transcribe for free (see Limites below)

With GEMINI_API_KEY set (from step 2):

gtw sample.mp4

Or pass the key directly:

gtw --gemini-api-key YOUR_API_KEY sample.mp4
gtw --gemini-api-key YOUR_API_KEY *.mp4

Done: sample.speakers.srt, sample.srt, sample.txt appear next to the input. Any local video/audio works (.mp4, .mp3, etc.).

Limits

This wrapper exists to get past the Gemini free tier limits:

• Max 2 API calls per minute: chunks are transcribed sequentially with a built-in delay (default 30s) between API calls.
• Max 30 minutes of audio per call: long files are split into 29m50s chunks (packed to the max so the 25 calls/day budget is used efficiently), transcribed, then merged.
• Max 25 API calls per day: once the daily quota is used up, the API returns HTTP 429. If you hit it, the tool prints the limits above: free tier users should try again tomorrow, or switch to a paid tier (enable billing) to keep going immediately.

Usage tracking: every API call is counted per day (PST, resets at midnight Pacific) and saved to ~/.cache/gemini-transcribe-wrapper/usage.json. Every run (gtw -v, gtw --help, or after transcription) prints today's count on the last line, e.g. 2026-08-28T220317-08:00 (PST) API calls today: 3/25 (free tier limit: 25).

Chunked transcription keeps the same global timeline as a single pass (word timestamps match within 0.5s; verified by verify-chunk-secs.sh). If you want an ffsubsync-aligned SRT for manual comparison, pass --ffsubsync-srt to also write <base>.ffsubsync.srt (aligned to the full audio via uvx --python 3.13 ffsubsync <audio> -i <srt> --max-offset-seconds=120 --gss --overwrite-input); the main .srt/.speakers.srt always keep the raw transcript timestamps.

.speakers.srt (speaker diarization, still a standard .srt for media players), .srt (no speaker labels), and .txt (clean text) are all generated at once by default. Don't need one? Turn it off:

gtw sample.mp4 --no-speakers-srt --no-txt   # keep only .srt
gtw sample.mp4 --no-srt                     # keep .speakers.srt + .txt

See gtw --help for all options (speaker name mapping, transcript re-rendering, temp dir, etc.).

License

MIT