geass 0.2.0

Opinionated CLI for generating transcripts to use with LLMs

Geass

CLI tool for audio transcription using faster-whisper. Supports local and remote transcription (via Modal), caching, and multiple output formats.

Features

• Transcribes audio files using faster-whisper models (e.g., base, large-v3).
• Local transcription with CPU/GPU support or remote via Modal with GPU options (T4, A100, etc.).
• Output formats: plain text, JSON, SRT.
• Aggregates segments into larger chunks with --interval.

Installation

pip install geass

Requires Python 3.10+. For local GPU use, ensure CUDA and PyTorch are set up.

Usage

List Available Models

geass list-models

Shows all faster-whisper models (e.g., tiny, large-v3-turbo).

Transcribe Audio

geass transcribe audio.wav

Transcribes audio.wav with default settings (base model, text output).

Options

• -m, --model: Specify model (e.g., -m large-v3).
• -f, --format: Output format (text, json, srt). Default: text.
• -s, --save: Save transcription to database.
• -r, --remote: Use Modal for remote transcription.
• -g, --gpu: GPU type for remote (e.g., -g A100). Default: A100.
• -i, --interval: Aggregate segments into intervals (e.g., -i 10 for 10s chunks).

Examples

Transcribe with a specific model and JSON output:

geass transcribe audio.mp3 -m large-v2 -f json

Remote transcription with saving:

geass transcribe audio.wav -r -s

Multiple files with SRT output:

geass transcribe file1.wav file2.mp3 -f srt

Database

Transcriptions are cached in ~/.geass.db. Tables:

• transcripts: Stores file metadata (path, duration, timings).
• segments: Stores transcription segments (start, end, text).
• cache: Maps file hashes to transcript IDs.
• defaults: Stores settings (model, etc.).

Notes

• Local GPU transcription uses float16 on CUDA, int8 on CPU.
• Remote transcription requires a Modal account and setup.
• The output is wrapped in XML <transcript> tags.