hehehai/AudioSmith

A promising all-in-one local voice studio for Apple silicon owners who want TTS, dubbing and transcription in one app without cloud accounts. The catch is a narrow hardware window and model downloads that scale with what you want to do.

The problem

You have a script, a book chapter or a video and want narrated audio in a chosen voice, perhaps with different characters and matching loudness, plus a transcript with speakers labelled. The usual route is to stitch together separate TTS, ASR, diarization and audio-editing tools, often web services, and move files between them. AudioSmith folds those steps into one Mac application: you import text or media, select a voice or model, and get a finished take, an M4B chapter, a dub, or a speaker-labelled transcript. The workflow changes from assembling pipelines to working inside a single studio window.

What's different

The core mechanism is running multiple speech models through MLX on the Mac's GPU, so speech generation and recognition happen locally once the models are downloaded, with no account required for the engine itself and offline operation after setup. On top of the models, the app adds workflow structure aimed at longer projects: script markup for pauses, pace and pronunciation, a story mode that casts Name: line dialogue and merges consecutive lines, audiobook chapters with ACX-style loudness, a four-step dubbing flow that can pull videos via yt-dlp, and a canvas for chaining steps over batches of inputs. That combination targets people producing finished audio, not just testing a model.

The trade-off is fragmentation of hardware and model licensing. Different capabilities need different model packs: 8 GB runs the Fast pack, 16 GB is recommended and required for one dialogue model, and 24 GB is needed for high-fidelity cloning, voice design, dubbing and the writing assistant. The models keep their own licences, and OmniVoice is explicitly non-commercial, so a produced voice may carry restrictions beyond the app's AGPL-3.0 licence. The dubbing route also involves optional downloads of ffmpeg and yt-dlp, so not every path is purely self-contained.

Who should try it

A good fit is an Apple silicon user with 16–24 GB of memory who already writes scripts, narrates books or dubs short videos and wants local processing with no per-minute cloud billing. The story and audiobook features are most useful when you have multi-character material rather than a single paragraph of speech. A poor fit is anyone on an Intel Mac, an older macOS version or 8 GB who expects high-fidelity cloning or dubbing, because those capabilities are tied to larger model packs and more memory. People who need Windows or Linux support, or who require commercially unrestricted voice models, should check the model licences and platform limits before investing time.

Requirements and limitations

The hard boundary is macOS 15 or later on Apple silicon (M1 or newer). Memory determines which features are available, not just performance, and disk use ranges from 1.3 GB to 17 GB for models, with dialogue models adding several more gigabytes. The download requires a free account on the vendor's site; the documented alternative is building from source, which changes the setup burden for people who want to avoid registration. The current release is a beta, with a changelog listing versions. The app is signed and notarized, but ongoing care means tracking model-pack requirements and the separate model licences as you move projects from experiment to publication.

HotRepo's take

The repository was 14 days old with 107 total stars when it appeared at #23 on HotRepo's NEW-REPOSITORY board (新项目榜, not the overall Trending board) on 2026-10-27. That is a single dated snapshot, so persistence cannot be judged from it. For an Apple silicon user with enough memory, the concrete thing to verify is whether the model pack you need matches your RAM and whether the licence of each voice model allows your intended use, especially OmniVoice's non-commercial terms.

Project source and documentation