Piper Transcription downloads

These archives are what the Piper desktop app downloads when a reader turns on Transcription in Settings, AI & Models. Piper verifies every archive against a sha256 compiled into the app before extracting it, and it never loads files from this repository any other way.

What is here

• transcription-engine-<platform>-<version>.tar.gz: a relocatable CPython from python-build-standalone with WhisperX (BSD-2-Clause), faster-whisper (MIT), pyannote.audio (MIT), and their dependencies, plus LGPL builds of FFmpeg in ffmpeg/ (see ffmpeg/LICENSE.txt and ffmpeg/SOURCE.txt). Each bundled package keeps its own license, recorded in its metadata inside the archive.
• transcription-models-<version>.tar.gz: Whisper medium.en converted by Systran (MIT, OpenAI's weights), the wav2vec2-base-960h alignment model (Apache 2.0, Meta AI), and NLTK's punkt_tab data (Apache 2.0).
• transcription-speakers-<version>.tar.gz: pyannote's speaker-diarization-community-1 (CC BY 4.0), downloaded only when a reader adds speaker identification.
• transcription-speakers-coreml-<version>.tar.gz: the CoreML speaker model from avencera/speakrs-models (CC BY 4.0), for Macs with Apple silicon.

piper-models.json in each Model archive records every file's source repository, pinned revision, license, size, and sha256.