Credits & acknowledgments

Built with remarkable open work.

Speak Easy Privately is its own product, privacy architecture, and macOS experience. Its local speech and writing capabilities are made possible by researchers and open-source contributors who chose to share their work. We are grateful to them.

Why these credits matter

Local-first products are rarely built alone.

The app can recognize speech and improve punctuation without a transcription server because model creators, researchers, and runtime engineers published software and model releases under licenses that permit others to study, integrate, and distribute their work.

Crediting that foundation is more than a license obligation. It is an honest description of how the product works and a thank-you to the people whose work made private, on-device dictation achievable for a small independent application.

CC BY 4.0

NVIDIA Parakeet TDT 0.6B v3

Parakeet is the speech-recognition foundation used to turn audio into multilingual text locally. The original model was created by NVIDIA. FluidInference published the Core ML conversion used by Speak Easy Privately so the model can run efficiently on Mac.

Speak Easy Privately pins and verifies the approved conversion. It packages or downloads the runtime files without changing the learned model weights.

Original NVIDIA model FluidInference Core ML conversion CC BY 4.0 license

Optional local writing models

Apache 2.0

Qwen3 GGUF

Created by the Qwen Team at Alibaba Cloud. Speak Easy Privately can use selected Qwen3 models for conservative, multilingual punctuation and capitalization while keeping the text on the Mac.

Qwen on Hugging Face

MIT

SAGE FredT5 Distilled 95M

Created by AI Forever, with the INT8 ONNX conversion published by krut42. It provides a compact specialist option for Russian punctuation and capitalization.

Original SAGE model INT8 ONNX conversion

The software that runs the models locally

Apache 2.0

FluidAudio

FluidInference’s Swift package for efficient local audio and speech-model inference.

View FluidAudio

MIT

llama.cpp

The ggml community’s high-performance local inference runtime used for Qwen GGUF models.

View llama.cpp

MIT

ONNX Runtime

Microsoft and contributors’ cross-platform inference engine used by the local SAGE pipeline.

View ONNX Runtime

What Speak Easy Privately adds

These upstream projects provide essential building blocks. Speak Easy Privately combines them into a cohesive macOS product with memory-only audio capture, language routing, conservative text projection, model integrity checks, sandboxed processing, encrypted optional history, permissions guidance, and insertion into the user’s active workflow.

The full edition packages the approved local components for an offline first run. Lite downloads only models the user chooses. In both editions, model versions and runtime files are pinned and checked before activation.

Licenses, trademarks, and respect

Speak Easy Privately preserves the applicable license texts and third-party notices inside every release. Project names and trademarks remain the property of their respective owners. Inclusion does not imply that NVIDIA, FluidInference, the Qwen Team, Alibaba Cloud, AI Forever, Microsoft, ggml contributors, or any other upstream creator endorses Speak Easy Privately.

Exact model revisions, conversion provenance, integrity hashes, and complete license documents remain available inside the app under Settings → Credits so they travel with the software even when the Mac is offline.

Found an attribution that should be clearer or more complete?