smrze is a work-in-progress CLI for creating local-only transcripts and summaries from media or transcript files.

The project is still early. Commands, setup, output shape, and model requirements may change as the tool is developed.

• Generate diarized transcripts from local media files
• Download remote media URLs with yt-dlp and process them locally
• Convert and decode media with ffmpeg
• Run transcription and speaker diarization locally
• Print results to stdout or write files into an output directory
• Produce transcript output as plain text, diarized text, speaker turn JSON, word timing JSON, or timestamped word text
• Summarize existing transcript files, media files, or remote media URLs
• Use Apple Foundation Models for summaries by default
• Fall back to Gemma 4 E2B when Apple Foundation Models refuse a transcript
• Select Gemma 4 E2B or Gemma 4 E4B explicitly for summaries
• Cache downloaded audio, transcripts, summaries, models, and MLX runtime assets under the app cache directory

Generate a transcript from a local media file:

smrze transcript ./meeting.mp4

Generate a transcript from a URL and write transcript.txt into an output directory:

smrze transcript "https://example.com/video" --output ./out

Generate a plain transcript without timestamps or speaker diarization:

smrze transcript ./meeting.wav --no-timestamps

Write word-level timing output as JSON:

smrze transcript ./meeting.wav --mode word --format json

Summarize an existing transcript:

smrze summarize ./transcript.txt

Summarize a URL with an explicit Gemma backend:

smrze summarize "https://example.com/video" --summary-backend gemma4-e2b

Use short aliases:

smrze t ./meeting.mp4
smrze s ./transcript.txt

Without --output, commands print their result to stdout.

With --output DIR, smrze transcript writes one of:

• transcript.txt
• turns.json
• words.json
• words.txt

With --output DIR, smrze summarize writes:

• summary.md

Use --open with --output to open the written file after creation.

• Rust toolchain
• ffmpeg for media conversion and audio fallback decoding
• yt-dlp for remote URL inputs
• macOS with Xcode for Apple Foundation Models and MLX-backed summaries
• SwiftPM network access on the first macOS build so it can fetch mlx-swift
• Metal Toolchain for MLX Gemma support

Install the Metal Toolchain before using MLX-backed Gemma summaries:

xcodebuild -downloadComponent MetalToolchain

cargo build

just fmt
just clippy
just test