AudioSmith

A voice studio that runs on your Mac.
Speech, voices, stories, audiobooks, dubbing, transcription and dictation, on Apple silicon.

Website · Download · Changelog

English · 简体中文

AudioSmith turns text into speech, clones and designs voices, gives a story a cast, reads a book into an M4B, dubs a video into another language, transcribes a recording with who said what, and types what you say into any app. Every model runs on your Mac's GPU with MLX: no account, no cloud service, no Python. It runs on your Mac, and once its models are downloaded it works offline.

Features · Install · Models · Privacy · Build from source · Licence

The current version, 0.4.0-beta.7, is a beta. Releases are signed with a Developer ID, notarized by Apple and update themselves. Expect rough edges; see Known limits and the changelog.

Features

Studio

• Write or import a script (TXT, MD, DOCX, EPUB, SRT, VTT). Markup adds pauses ([pause 400ms]), pace ([slow], [fast]), emphasis, spelling and one-off pronunciations ([[word|say]]).
• Pick a model by what it is for: Standard voices, Fast clone, High-fidelity clone, Voice design, Fast English voices, English clone or Multilingual all-rounder. The technical name is one click away.
• Few controls: the model, one Voice list and the speed. Everything else is under Advanced, filled with settings measured for each model.
• Long scripts render in chunks and play while they render. Text you did not change comes from a cache.
• Export audio: WAV or M4A, and MP3, FLAC or Opus with ffmpeg. Even out the loudness to −16, −18 or −23 LUFS, trim the silence, add tags, save the script and subtitles beside it.

Voices

| Discover voices | My voices | |---|---| | | |

• Discover 37 chosen voices, mostly Chinese and English: designed voices whose reference ships with the app, ready to speak at once, and the best built-in speakers. Preview them and save the ones you like.
• Clone a voice from a recording or from any audio or video file. AudioSmith picks the best stretch and transcribes it. A recorded consent statement marks a voice as your own.
• Checked references: every reference is rated (good, fine, needs work) with a tip or two, and cleaned: noise removed only when there is some, silence trimmed, loudness evened out. The original is kept.
• Design a voice from a sentence in English or Chinese ("a warm narrator in her thirties, calm and clear"), or start from one of eight presets.
• Hear before you pick: every voice list (Studio, Stories, Audiobook, Dub) has a ▶ on each voice: built-in speakers play a short bundled sample, even before their model is downloaded; your own voices play their sample.
• Portable voices: a model's built-in speaker or a designed voice can be turned into a clone that every cloning model can speak with. It sounds close to the original, not the same, and says so.
• Deleting a voice says where it is used, and Undo in the toast brings it back for about 10 seconds.
• Voices are .asvoice files you can share. VoiceStudio's .ovsvoice files import too.

Stories

• Write Name: line and AudioSmith casts the characters. Stage directions such as MOM (softly): become the line's direction instead of being read out.
• Dialogue that flows: the gaps follow the conversation (Dialogue, Reading or Relaxed). A character's consecutive lines are said in one go, and each voice is brought to one loudness.
• AI director: the writing assistant (Qwen3 4B, on this Mac) reads the whole story and gives each line one of 13 emotions with a strength, kept coherent across the scene; you can change any of them. Models that take instructions hear them; the others read the lines as written, and the page says so.
• Whole dialogue: a dialogue model speaks the characters' lines a scene at a time, so each line answers the one before it. Dialogue · Light (SoulX-Podcast) needs 16 GB of memory, Dialogue · One take (MOSS-TTSD) 24 GB. Narration stays on the story's model.
• Export a WAV, an M4B, or one stem per character.

Audiobook

• Import an EPUB, TXT, Markdown or DOCX book. Chapters come from its headings.
• Dialogue passages: mark a scene with [dialogue] … [/dialogue] ([对话] in Chinese) and one Name: line per turn, and a dialogue model speaks it with the characters answering each other. Without the model the passage is read line by line.
• Loudness per chapter for ACX (Audible) or podcasts. A read-along player follows the text.
• Export an M4B with chapters, cover and tags, a cue sheet, or one file per chapter.
• Resume after quitting. After an edit, only the changed lines are rendered again.

Dub

• Drop in a video or audio file, or paste a link (yt-dlp, which you turn on). Before anything downloads, the page shows the cover, the site, the title and the length, then the download's real progress. New dubs are made in Chinese; when the video is already Chinese they switch to English. Add more languages as you like.
• Four steps:

• Conversations recorded together: for dubs into Chinese or English with two speakers or more, Dialogue · Light records each conversation as a whole, so answers sound like answers, then fits every line to its time. On by default for new dubs; needs 16 GB of memory.
• Edit lines on a timeline (pinch or ⌘ with the wheel to zoom, swipe to pan), compare A/B, preview the video, and re-dub only what changed. A 2-hour video with 2,000 lines stays quick.
• Export an MP4 with one audio track per language and soft subtitles, or M4A, WAV, SRT, VTT, ASS and stems. Batch dubs a whole folder, and can watch one.

Transcribe and dictation

• Transcribe recordings and videos with word timings. Qwen3-ASR handles Chinese, English and 28 more languages; Whisper covers the rest and translates speech into English.
• Transcription with speakers says who says what (S1, S2…) in one pass; click a speaker to name it. Copy, TXT, SRT, VTT and JSON exports carry the names.
• Dictation anywhere: press ⌘⇧Space in any app, speak, and the text is typed where your cursor is, about 0.2–0.3 s after you stop. Your clipboard is kept. A menu bar icon, a small pill and sounds you can choose show when it listens.

Workflows and batch

• Workflows: chain steps on a canvas (sources, speak, transcribe, translate, convert, even out, condition, export), start from a template, and run them over up to 50 inputs. A run continues after a restart.
• Batch queues dubs and watches a folder.

For agents and scripts

• OpenAI-compatible API on http://127.0.0.1:3901/v1: speech, transcriptions, translations, models and voices. Point any OpenAI SDK at it.
• AudioSmith's own API (/api) drives every page: voices, stories, audiobooks, dubs, batch, workflows, exports. Jobs report progress over SSE, and /ws/tts streams speech.
• MCP server over HTTP (/mcp) and stdio (AudioSmith --mcp-stdio), with 33 tools. Deleting needs confirm: true; cloning needs consent and works only from this Mac (or with the API key).
• Integrations page: one prompt sets up Claude Code, Codex or Cursor (server and agent skill); there are snippets for n8n, curl, Python, JavaScript and the OpenAI Agents SDK.

curl http://127.0.0.1:3901/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{"model": "tts-1", "voice": "alloy", "input": "Build finished. All tests passed."}' \
  -o speech.wav

The full reference is docs/local-api.md, and /docs in the running app.

And more

• Tools: compare voices A/B, speak a recording again in another voice, fit text to a duration, probe a file, check a watermark.
• Invisible watermark: generated audio carries an AudioSeal mark, on by default (Settings → Privacy), and Tools can detect it.
• Calls: an AI phone agent on your own Twilio account. It always says it is an AI, and uses only designed or consented voices. Off until you set it up.
• The app: a sidebar with search (⌘F) and sections that fold; what is running shows as cards at its foot, and the Tasks page lists every job. Projects and History libraries, one job queue that resumes after a restart, the command palette (⌘K), ⌘1–9 for the main pages, light and dark themes, English and 简体中文.
• Memory: models you have not used for 30 minutes are unloaded (Settings → Models & network → Memory) and load again when needed.

Install

Requirements

• macOS 15 or later, on Apple silicon (M1 or later).
• Memory: 8 GB runs the Fast pack; 16 GB is recommended, and needed for Dialogue · Light; 24 GB runs the Quality pack (high-fidelity cloning, voice design, dubbing and the writing assistant) and Dialogue · One take.
• Disk: 1.3–17 GB for models, depending on the pack; the dialogue models add 2.4 GB and 10.9 GB.

1. Download the DMG from actnow.dev/download/audiosmith, AudioSmith's page on our website. AudioSmith is free; the download needs a free account. Rather not sign up? Build it from source. The current version is a beta; the changelog lists every version.
2. Drag AudioSmith to Applications and open it. Releases are signed and notarized, so Gatekeeper lets them open; macOS only asks once to confirm an app from the internet.
3. The setup wizard checks your Mac (macOS, chip, memory, disk, network), suggests a model pack, and asks about privacy and dictation. You can skip the pack and download models later on the Models page.

Updates. Once a day AudioSmith checks actnow.dev for a new version (no account needed). A new one shows as "Update available" at the foot of the sidebar, with its notes; it installs when you say so. Updates are signed and checked before they install, and usually download only what changed. AudioSmith → Check for Updates… checks at once; Settings → About → Software updates has the other choices.

Beta updates. Settings → About → Software updates → Receive beta updates. It is on in a beta (such as 0.4.0-beta.7) and off in a release until you choose. Turning it off never goes back to an older version: a beta stays until a newer release is out.

Downloads come from the official sites or their mirrors, whichever is faster, through the system proxy: models from huggingface.co or hf-mirror.com, ffmpeg and yt-dlp from github.com or mirror.actnow.dev (ActNow's mirror), each checked before use. In mainland China, pick Mirrors only under Settings → Models & network → Download from if downloads stall.

Models are kept in ~/Library/Application Support/ActNow/models, shared with the other ActNow apps, and left out of Time Machine (Settings → Advanced turns backups back on).

Permissions are asked only when a feature needs them: the Microphone for recording a voice, dictation and live transcription, and Accessibility for dictation to type into other apps.

Models

AudioSmith names models for what they do. Their technical names, sizes and licences are in each model's Details on the Models page.

| In the app | Model | What it does | Licence | |---|---|---|---| | Standard voices | Qwen3-TTS 0.6B CustomVoice | built-in Chinese and English speakers, with directions | Apache-2.0 | | Fast clone | Qwen3-TTS 0.6B Base | quick voice cloning | Apache-2.0 | | High-fidelity clone | Qwen3-TTS 1.7B Base | the most faithful cloning | Apache-2.0 | | Voice design | Qwen3-TTS 1.7B VoiceDesign | voices from a description | Apache-2.0 | | Fast English voices | Kokoro 82M | natural English voices; light | Apache-2.0 | | English clone | Chatterbox Turbo | fast English cloning | MIT | | Multilingual all-rounder | OmniVoice | 600+ languages, cloning, design, exact durations | CC BY-NC 4.0: non-commercial; asks you to accept it first | | Chinese & English transcription | Qwen3-ASR 0.6B | transcription and dictation, 30 languages | Apache-2.0 | | Multilingual transcription | Whisper large-v3 turbo | other languages; speech into English | MIT | | Transcription with speakers | MOSS-Transcribe-Diarize 0.9B | who says what, in one pass | Apache-2.0 | | Dialogue · Light | SoulX-Podcast 1.7B | conversations recorded together, up to 4 speakers; 2.4 GB, 16 GB memory | Apache-2.0 | | Dialogue · One take | MOSS-TTSD v1.0 | a scene as one recording, up to 5 speakers; 10.9 GB, 24 GB memory | Apache-2.0 | | Writing assistant | Qwen3 4B Instruct 2507 | translation, clean-up, the AI director | Apache-2.0 | | Word timing · Speech detection · Noise removal | Qwen3-ForcedAligner 0.6B · Silero VAD · DeepFilterNet 3 | | Apache-2.0 · MIT · MIT/Apache-2.0 | | Fast voice separation · Speaker turns | HTDemucs · Sortformer (fallback, ≤ 4 speakers) | | MIT · NVIDIA Open Model License | | Invisible watermark | AudioSeal | | MIT |

Measured on a MacBook Air M4 with 24 GB:

| What | Result | |---|---| | Dictation (Qwen3-ASR) | 0.2–0.3 s after you stop speaking | | High-fidelity clone | about real time: 940 Chinese characters → 3.5 min of audio in about 3¼ min | | Standard voices | about 1.7× real time | | Transcription with speakers | a 10-minute recording in about 4 minutes | | Dialogue · Light / One take | about 0.7× / 0.8× real time | | Writing assistant | 36–40 tokens/s |

Privacy

• Speech is made and recognised on your Mac. The engine never connects to the network; it loads the models the app downloaded.
• AudioSmith connects only for: model downloads (Hugging Face or its mirror), ffmpeg and yt-dlp when you download them, videos (and their covers) from links you paste, a cloud writing or translation provider only if you set one up, and the daily update check (nothing about you is sent).
• No analytics: no usage data, crash reports or identifiers, and no setting to turn them on. The app's log stays on this Mac, masks keys and tokens, and never holds your scripts, transcripts or prompts.
• The Local API listens on 127.0.0.1 unless you share it on your network (with a PIN).
• Keep text on this Mac: one switch limits writing help and translation to local models.
• Your library is in ~/Library/Application Support/AudioSmith (it can move to another disk). Keys live in the login Keychain.

Build from source

You need Xcode 26 (Swift 6.2 or later) and Go 1.27.3 (GOTOOLCHAIN=auto downloads it).

scripts/fetch-engine-deps.sh    # upstream mlx-audio-swift + AudioSmith's patches
scripts/build-engine.sh         # the Swift engine (xcodebuild; about 4 min the first time)
go tool mygo dev                # run with live reload
go tool mygo build              # build/darwin-arm64/AudioSmith.app and a DMG

• The UI is Go with MyGo, in app/ (go run ./app runs it without bundling); the files it embeds are in assets/. The models run in a Swift helper, audiosmith-engine, on mlx-audio-swift and mlx-swift.
• CGO_ENABLED=0 go test ./... runs the whole UI headless on any OS, with the engine faked; scripts/mac.sh builds and tests on a Mac from Linux.
• More: architecture, engine protocol, Local API and MCP, releasing.

Known limits

• Beta. 0.4.0 is not final yet; expect changes between betas.
• ffmpeg is optional and not bundled. Without it AudioSmith reads and writes WAV, AIFF, M4A, MOV and MP4, and reads MP3. With it (Homebrew's, or one you choose) it also exports MP3, Opus and FLAC, reads WebM and MKV, and burns subtitles into videos.
• Dictation types the text when you stop, not word by word. The first dictation after a long pause can take about a second longer while its model loads.
• Dialogue models are evaluated for Chinese and English only; other languages are read line by line. Dub's conversations use Dialogue · Light only, and Dialogue · One take needs 24 GB of memory.
• Links that need a sign-in (such as X or Douyin) need a cookies.txt export from your browser.
• The interface is in English and Simplified Chinese.
• The watermark does not survive time-stretching beyond about ±5 %, so dubs are marked after fitting.
• Calls need your own Twilio account and a public tunnel.
• Builds you make yourself are ad-hoc signed: they don't update themselves, macOS forgets their Accessibility permission after each rebuild, and dictation plays the system sounds (the ActNow sound library ships only with releases).

Contributing

Bug reports, ideas and pull requests are welcome: see CONTRIBUTING.md for the checks to run. Before we can merge your first pull request, you sign the Contributor License Agreement by commenting on it; a bot asks you.

Licence and credits

Copyright (C) 2026 hehehai and the AudioSmith contributors.

AudioSmith is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, version 3 (AGPL-3.0). It is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; see the licence for details. If you run a modified AudioSmith for others to use over a network (its Local API, for instance), the AGPL requires you to offer those users the source code of your version. Versions up to and including 0.4.0-beta.1 were released under the Apache License 2.0 and stay under it.

Model weights keep their own licences (see Models); OmniVoice's is non-commercial. Third-party material and libraries are listed in NOTICE, with their full licence texts in THIRD-PARTY-NOTICES.txt (the app carries both).

• Models by the Qwen team (Qwen3-TTS, Qwen3-ASR, Qwen3 4B), hexgrad (Kokoro), Resemble AI (Chatterbox), k2-fsa (OmniVoice), Soul AI Lab (SoulX-Podcast), OpenMOSS (MOSS-TTSD, MOSS-Transcribe-Diarize), OpenAI (Whisper), Meta (HTDemucs, AudioSeal), NVIDIA (Sortformer), Silero and DeepFilterNet. MLX conversions by mlx-community and others.
• Runs on MLX, mlx-swift and mlx-audio-swift.
• The interface is built with MyGo (MIT).
• Icons: Solar by 480 Design, CC BY 4.0 (see NOTICE).