PolyScribe
PolyScribe is a fast, local-first CLI for transcribing audio/video and merging existing JSON transcripts. It uses whisper-rs under the hood, can discover and download Whisper models automatically, and supports CPU and optional GPU backends (CUDA, ROCm/HIP, Vulkan).
Key features
- Transcribe audio and common video files using ffmpeg for audio extraction.
- Merge multiple JSON transcripts, or merge and also keep per-file outputs.
- Model management: interactive downloader and non-interactive updater with hash verification.
- GPU backend selection at runtime; auto-detects available accelerators.
- Clean outputs (JSON and SRT), speaker naming prompts, and useful logging controls.
Quickstart
- Install Rust (rustup) and ffmpeg, then build:
- rustup install stable
- rustup default stable
- cargo build --release
- Download a model (first run can prompt you):
- ./target/release/polyscribe --download-models
- Transcribe a file:
- ./target/release/polyscribe -v -o output my_audio.mp3 This writes JSON and SRT into the output directory with a date prefix.
Model locations
- Development (debug builds): ./models next to the project.
- Packaged/release builds: $XDG_DATA_HOME/polyscribe/models or ~/.local/share/polyscribe/models.
- Override via env var: POLYSCRIBE_MODELS_DIR=/path/to/models.
- Force a specific model file via env var: WHISPER_MODEL=/path/to/model.bin.
Most-used CLI flags
- -o, --output FILE_OR_DIR: Output path base (date prefix added). If omitted, JSON prints to stdout.
- -m, --merge: Merge all inputs into one output; otherwise one output per input.
- --merge-and-separate: Write both merged output and separate per-input outputs (requires -o dir).
- --set-speaker-names: Prompt for a speaker label per input file.
- --update-models: Verify/update local models by size/hash against the upstream manifest.
- --download-models: Interactive model list + multi-select download.
- --language LANG: Language code hint (e.g., en, de). English-only models reject non-en hints.
- --gpu-backend [auto|cpu|cuda|hip|vulkan]: Select backend (auto by default).
- --gpu-layers N: Offload N layers to GPU when supported.
- -v/--verbose (repeatable): Increase log verbosity. -vv shows very detailed logs.
- -q/--quiet: Suppress non-error logs (stderr); does not silence stdout results.
- --no-interaction: Never prompt; suitable for CI.
Minimal usage examples
- Transcribe an audio file to JSON/SRT:
- ./target/release/polyscribe -o output samples/podcast_clip.mp3
- Merge multiple transcripts into one:
- ./target/release/polyscribe -m -o output merged input/a.json input/b.json
- Update local models non-interactively (good for CI):
- ./target/release/polyscribe --update-models --no-interaction -q
Running tests and tools
- cargo test
- cargo clippy --all-targets -- -D warnings
- cargo build (preferably without warnings)
Model downloader
- Interactive: ./target/release/polyscribe --download-models
- Non-interactive: relies on defaults; set WHISPER_MODEL or POLYSCRIBE_MODELS_DIR when needed.
Documentation index
- docs/usage.md – complete CLI reference and workflows
- docs/development.md – build, run, and contribute locally
- docs/design.md – architecture overview and decisions
- docs/release-packaging.md – packaging notes for distributions
- docs/faq.md – common issues and solutions
- docs/ci.md – minimal CI checklist and job outline
- CONTRIBUTING.md – PR checklist and workflow
CI status: [CI badge placeholder]
Examples See the examples/ directory for copy-paste scripts:
- examples/transcribe_file.sh
- examples/update_models.sh
- examples/download_models_interactive.sh