最近更新

0.5 changes how you work in AudioSmith, not how it looks: colours, type and buttons are the same. You start from one box on Home; every tool starts its work at the top of its own page and lists what it made below; results hand off to the next tool; everything you make is in Assets and every long piece of work in Projects.

Where things are now

Before	Now
Studio	Text to speech
History	Assets
Voice gallery, My voices	Voices — Discover, Mine
Design a voice, Clone a voice	New voice — Design a voice, Quick clone
Library	Projects (search in it with the sidebar's search or ⌘K)
Models, Engine, Logs	Settings → Models & network, Engine, Logs
Integrations, Settings → Integrations, Local API	Developers — Overview, Local API (Settings → Integrations points there)
Tools	Developers → Tools
Activity cards at the sidebar's foot	the bell beside Settings
Top bar and breadcrumbs	each page's own header
Stories' and Audiobook's Project menu	the page's list, and the editor's ⋯

Old links (audiosmith://gallery, /history, /tools, /integrations…) open the new places; the Local API and MCP are unchanged. ⌘1…⌘8 are Home, Text to speech, Voices, Projects, Dub, Speech to text, Stories, Audiobook; ⌘F puts the keyboard in the sidebar's search.

Start from Home

• AudioSmith opens on Home: 「今天想做点什么？」 and a box. Type, paste a link or drop a file and press Enter (Shift-Enter for a new line): text opens in Text to speech as the script, a video or a link starts a dub, a recording is transcribed, a text file or an ebook becomes an audiobook. A row of tools and your recent work are under it, and five samples. With the writing help's model loaded, 「智能分派」 lets it pick between Text to speech, Stories and Audiobook.
• The sidebar lists places — Home, Voices, Projects, Assets — then the tools: Text to speech, New voice, Stories, Audiobook, Dub, Speech to text, Voice isolator, Dictation and ··· More (Workflows, Batch dubbing, Calls). The search is at its top (pages, actions, voices, projects, samples); Developers, Settings and the bell (tasks and notifications) are at its foot.
• Every page has its own header: its title, how many things it lists, and its own tools with at most one main button, all of one height. ⌘[ ⌘] and the mouse's back and forward buttons go through the pages you saw; an item's page shows its list's name before its title, and a click on it goes back to the list. In a narrow window the search becomes a magnifier and the less used tools go into ⋯.
• ⌘K searches everything. Empty, it offers quick actions and your latest work; typing groups pages, actions, voices (Enter uses one in Text to speech), projects, samples and takes by their script, and always offers to read what you typed in Text to speech.

Every tool starts on its page

• Text to speech: the script fills the left; chips under an empty script (Tell a story, Product narration, Podcast intro…) fill it with an example in the voice's language. Generate (⌘↩) is at its foot with the latest take. On the right, Settings (voice and model pickers, 「✦ Recommend」, speed, language, Advanced) and History (this page's takes, with Export, Reuse settings, Regenerate with another voice, Save as project, Delete).
• Dub: drop a video or a recording, or paste a link, pick the language and press 「配音」. It starts at once and shows on top of the list below with its progress; a click opens the dub's editor.
• Speech to text: drop, upload, link or record, choose the language, speakers and Fast or Accurate, and send. A transcript opens on its own page and hands off to 「配音」 and 「在文字转语音里朗读」.
• New: Voice isolator. Drop a song, a video or a noisy recording and get the voices alone, the music and effects, and the voices denoised, on this Mac (HTDemucs, DeepFilterNet); each track plays, saves and goes on to Speech to text or Dub.
• Stories and Audiobook open on a composer over your stories or books: describe a story or paste a script, or paste or import a manuscript and pick the narrator, then 「新建」. The project opens in the editor you know; the list's name in the header (or ⌘[) saves it and goes back to the list. Save and Save a copy moved into the editor's ⋯.
• New: 「帮我写」 on Stories. With the writing help's model on this Mac, turn it on and describe a story in a sentence: the writing help drafts a short scene for two or three voices in your description's language — written, without filler words — then the story is made, cast and opened. A progress line shows while it writes, with Cancel.

Voices

• 「发现 | 我的」 switch in the header.
• Discover: menus for language, gender, age and chips for use; 「Recommended」 cards of one width, then 「All」. ▶ plays in the row; a click opens the voice's details at the right: a large preview, 「Try a line」, and 「Use in Text to speech」, 「Use as a story character」, 「Narrate an audiobook」, 「Save to My voices」.
• Mine lists your voices with filters (All, Cloned, Designed, Saved); ⋯ renames, duplicates, exports or deletes (with Undo). A saved voice's details hold what the profile editor had.
• New voice offers Quick clone (a guided page: add a recording → check → name and save → done) and Design a voice (a description, preset chips, three candidates to pick from).

Projects, Assets and the samples

• Projects lists stories, audiobooks, dubs, transcripts and voice isolations, newest first, with type, date, length and state; search (titles, and the scripts of stories, books and transcripts), filter by type, table or cards; ⋯ opens, renames, duplicates or deletes. Isolations show their file's name, not its path.
• Assets gathers every output and every file you brought in — takes, renders, exports, isolated tracks, the recordings and videos you used — filtered by type and by tool. Each row plays inline, opens in its tool, and sends on to Speech to text, Dub, Voice isolator or (its text) Text to speech.
• Samples — 两人对话的故事, 一章有声书, 一段产品旁白, 一段配音短片, 一段会议转写 — on Home's 示例 tab, at the top of Projects and in ⌘K. Opening one copies it into your library as a normal project you can edit or delete, and opens it; opening it again asks whether to open your copy or make a new one. Each opens ready to play, and no feature introduction comes first: 「雨夜修表铺」, a scene for two characters and a narrator (the dialogue spoken by the multi-speaker model); 「北站的末班车」, a first chapter with two characters; thirty seconds of narration for a made-up lamp, as Text to speech's latest take; 「天空为什么是蓝色的」, a short English explainer of our own dubbed into Chinese, open in the dub editor; and a release meeting for two, transcribed with its speakers. All their voices are built-in ones.
• A dub copied or moved to another folder (a sample opened, the data folder moved in Settings → Storage) keeps its lines: editing one renders that line again, not every line. Dubs made before render each line once more the next time they run.

Help where you need it

• The first visit to Dub, Speech to text, Voice isolator, New voice, Stories and Audiobook shows a short introduction: three points and 「开始」. Each shows once; Settings → General → 「重新显示功能介绍」 shows them again.
• Empty lists show ghost rows that say what will be there.

Lists, panels and controls

• Every list — voices, projects, assets, dubs, transcripts, isolations, Text to speech's history, Home's recent work — has the same rows: no column titles and no lines between rows, rounded rows that tint under the pointer, while open and while playing, a click opens the item, ▶ and ⋯ at the row's end, and small icons before the length, date, type, language and use (their tooltips name them). Delete on a selected row asks to delete it.
• Details and tool panels float at the right of the page, under its header, sliding in and out: a voice's or a take's details no longer narrow the list; Text to speech's, a dub step's and a story's or book's panel keep their room beside the page while it is wide enough and float over it in a narrow window. ✕ or Escape closes them; a click on another row shows that one.
• Header filters, buttons and switches are as tall as the search field; Developers' and Text to speech's tabs are switches like Voices'.
• Play and pause are new icons (Hugeicons) everywhere audio plays; a preview shows pause, not stop, while it plays.
• Buttons and menu items no longer end in 「…」: Export, Import, Delete, Set voice, Settings and the others are just their names. Text that is still loading or waiting keeps its 「…」.
• Badges in a row — language, gender and age, use, a voice's kind, a project's type, an asset's source — are filters: a click shows only those, a second click clears it. ▶, 「Use」 and ⋯ in a row light up under the pointer, and every voice in Discover and Mine has a preview.
• Pages use a little more of the window: 16 points at each side instead of 24, the same for the header and the page. Home stays a centred column.
• 「··· More」 in the sidebar folds open in place (Workflows, Batch dubbing, Calls) and reads 「Less」; it opens by itself when one of those pages is opened from elsewhere.
• The menu bar icon shows a red dot while AudioSmith is generating (speech, dubs, renders, transcripts, dictation, the writing help) and a violet one while an agent is connected over MCP; downloads and waiting work don't count.

Dub editor

• The steps — transcribe, translate, voices, generate, export — are icons in the page's header: done ones are tinted, the current one stands out, running ones spin, failed ones are red; the tooltip names the step and its state.
• No footer: the step's main action sits in the transport row (or the header before there is a transcript), the state and progress ring in the header beside the panel button.
• The timeline: click the ruler to seek, drag the playhead, click a block to select its line (double-click plays from it); dragging empty space no longer scrubs. The lane is lower, and zoom, replay and ±5 s share the transport row. Space plays and pauses, Home goes to the start, J and L step back and forward.

Developers

• Developers opens on 「接入你的 Agent」: choose 编程代理, MCP or cURL and copy the snippet — a prompt for your coding agent, the MCP server's configuration, or a first request — then see 「等待第一次请求…」 turn into the time of the last request. Under it, links to the API reference, the skill and the Local API's settings, and every app and agent with its setup. The Local API and the tools (A/B voices, Convert, Rate fit, Direction, File probe, Watermark) are its other tabs.
• Agents can do everything the app does (except dictation, phone calls, the microphone, accepting licences and changing settings): the MCP server has 57 tools — new ones isolate voices, list assets and samples and open a sample, rename, duplicate, export and import voices, preview any voice, check a recording before cloning, give three design candidates, draft a story with the writing help, say a recording again in another voice, download a model, detect the watermark, probe files, suggest a delivery, check whether a line fits, and read the defaults. Dubs start from a link, transcripts take Fast or Accurate. The skill describes every tool with short recipes. Tested with ChatGPT (Codex) and WorkBuddy.
• Agents ask before they delete or download. A request that names the action is not taken as a yes: the agent says what goes or how big the download is and waits; clients that support it show AudioSmith's own confirmation. Tools say which ones only read, which delete and which go online. Agents are told to use the tools, not AudioSmith's window.

Fixes

• Stories: in a script without blank lines, a narration line after 「Name: …」 is narration again, not that character's line.
• A voice copy is named in the app's language.

Dub timeline: zoom and pan

• Zoom the waveform timeline with a trackpad pinch, ⌘ or ⌥ with the scroll wheel, or ⌘= and ⌘−, around the pointer; ⌘0 or a double click on the zoom readout shows it all. It zooms from the whole source to 100 pixels a second.
• Pan with a sideways two-finger swipe (with the system's inertia), ⇧ with the wheel, a middle-button drag, or the new overview strip under the zoom controls, which shows the part in view over the whole source; drag it, or click beside it to jump there.
• While playing zoomed in, the view turns the page as the playhead reaches its edge, unless you panned in the last few seconds.
• Dense lines no longer crowd the timeline with numbers: a number shows only where it fits and leaves room after the last one.

Dub: see the video before it downloads

• Paste a link and the Dub page shows what it is within seconds, before anything downloads: the cover, the site (YouTube, Bilibili, Vimeo, TikTok and the others yt-dlp knows), the title, channel, length, picture size, estimated download size, date, views, the description (Expand shows all of it) and the caption languages. It is kept in the dub, so it shows again without the network.
• While the video downloads, a thin bar under the cover gives the size and time left; once it is downloaded, the centre shows the video itself, with play, while transcribing and the steps after it report in the footer.
• A link that cannot be read (a private or removed video, one that needs a sign-in, a site that is not supported) says so before any download, with the site's reason, and Try again. Cancel stops the reading as it stops a download.
• A dropped video or audio file shows its poster, length, picture and size the same way.
• The cover is fetched by the app itself, through the same proxy as the download, from public addresses only, at most 5 MB, and kept only if it is a JPEG, PNG or WebP picture.

Faster with long projects

• Dub: the editor builds only the lines in view; a 2-hour video's 2,000 lines no longer slow every frame (190 ms → under 1 ms a frame).
• Audiobook and Stories: a long manuscript (a novel) is checked and its draft saved once you pause typing, away from the window, not at every key; typing in 500,000 characters is as quick as in a short text.
• Studio and Audiobook editors: wrapping long text remembers what it measured, so opening, resizing and typing in a long manuscript no longer stall (rows still never pass the editor's edge).
• The invisible watermark is added while the audio streams through, so watermarking an hours-long dub no longer holds the whole recording in memory several times.
• The engine keeps up to 6 models loaded (small ones don't count) and lets the memory budget decide, so a story render no longer reloads a model at every line.
• Finished tasks older than 30 days are removed at startup; voice previews no longer copy their audio every frame; the render cache is tidied at most once a minute.

Security

• Voice names are cleaned wherever they come in (imported .asvoice and .ovsvoice files, the Local API, MCP, the app, Discover voices): line breaks and control characters become spaces, and a name keeps at most 80 characters. Names saved earlier are cleaned the next time they are saved.
• The agents' skill and setup prompt list voice names as data (quoted, cut short), not as text an agent could take for instructions.
• MCP voice cloning and design work only from this Mac (or with the API key), like the Local API's; over the LAN, a voice's audio is its sample, never the reference recording.
• Built with Go 1.27.3 (security fixes in Go's standard library).

Logs

• An operations log in the app's log (source “ops”, one key=value line each): creating, renaming, deleting, importing and exporting voices, projects, dubs, transcripts, workflows and takes, with who asked (the app, the Local API, LAN, MCP); settings changes by name; each job's start, end, failure (with its error) and cancel (with who cancelled and how long it ran); downloads; Local API and MCP calls that change something, with their status; errors shown as toasts; the app's start and each step of its quit. Never script text, transcripts, prompts or secrets.
• Logs and diagnostic bundles mask Local API keys, Twilio tokens, keys read from the Keychain and bearer tokens, and show the home folder as “~”; the engines' own log files are masked too.
• Errors that were dropped silently in background work (cancelling jobs, export records, notifications, files the app writes or removes…) are logged.

Fixes

• Studio and Audiobook script editors: the last character of a line was cut in half at the right edge on the Mac (a Chinese comma or full stop measured half as wide on its own as within its line). Lines now wrap where they are drawn.
• Importing a video from a link shows its real progress (percent, size, time left) on the page, in the footer and on the sidebar card, with the video's title; Stop on the card and Cancel in the footer really stop it (yt-dlp's helper process kept downloading before), and activity cards no longer show an empty bar.
• A copy built on the Mac says, in Settings → About → Software updates, why it cannot update itself, with a link to the download page.
• My voices: Delete… (and the other items) in a voice card's ⋯ menu did nothing — no dialog, no log line — unless the voice's editor was open: the menu went away with the pointer before the item chosen reached it.
• Deleting a voice says where it is used (stories, books, dubs, workflows, the Studio, batch dubbing, calls, agents) and can be undone from its toast for about 10 seconds; its files wait in the library's trash until then. Places that named it show “Voice deleted” to choose another one, a story or book that uses it stops with a clear message, and the Studio says its voice went — never another voice in silence.
• Quitting never hangs: edits not saved yet (drafts, the dub editor's changes) are written first, and jobs that do not stop when cancelled — such as a stuck download — are left after a few seconds (they continue or wait for Retry at the next start). An engine that stops reading its requests is restarted instead of blocking.
• A job after an interrupted one (the engine stopped twice in a stage) no longer waits forever: it fails, saying why, the dub or book is no longer shown as busy, and Resume continues it. Chains stuck this way in an earlier version recover at the start.
• The language model's turn on the GPU lasts until its request has really stopped, so the next job is not refused as busy.
• Models moved by a move that did not finish (cancelled, a model in use by another app, the app quit) no longer show as not downloaded; the move continues after a restart.
• Deleting a dub waits for its running stage to stop; two runs of the same dub started at once make one.
• Stories and Audiobook: a rename on the Projects page no longer marks unsaved changes as saved; saving over a project changed elsewhere (by an agent through the Local API) asks whether to save a copy or replace it, and the Local API refuses an edit made on an outdated read.
• Moving the data folder measures it and cleans up a failed copy in the background, with the dialog saying so, instead of freezing the window.
• Smaller fixes: a crash in background work is logged instead of closing the app; a draft that could not be written is tried again; unfinished temporary files are cleaned in every folder; downloads left running by a crash are stopped at the next start.

The window in the ActNow family's layout

• The sidebar starts with a search field (⌘F): it searches projects, voices, takes, transcripts and exports, and the Projects page shows what it finds ("3 found", Clear search). ⌘K opens the command palette from the button in it.
• Studio, Projects and History come first; Create (Stories, Audiobook, Dub, Transcribe), Voices and Automate are sections that fold. The ⌘1–⌘9 keys are as before.
• No status bar: what goes on shows as cards at the foot of the sidebar — a stopped engine (Restart), dictation and calls under way (Stop, Hang up), each story, book, dub and workflow being made, takes, exports, transcriptions and downloads, several of a kind as "2/5", a finished one for 3 seconds. The pointer over them opens the stack; View all opens the new Tasks page, with every job that runs or waits.
• Under the cards: Settings, the logs panel (⇧⌘L) and Tools, and the notifications' bell (a dot while some are unread).
• Models and Integrations are Settings sections now (their links open them); Engine and Logs left the sidebar (the palette, the Go menu, Settings → Advanced → Diagnostics). ⌘E's quick switch is a small dialog with the engine's state and the model in use.
• Settings: 17 sections become 11, each a column of groups with a line under every setting: General (with Appearance), Studio (with Pronunciation), Dubbing, Dictation (with Permissions), Writing help, Integrations (apps and agents, then the Local API), Audio tools, Models & network (the models, then downloads, storage, the Hugging Face token, memory and licences), Privacy, Advanced (storage, usage, diagnostics and the self-check) and About. Old links open the part they were.
• Page titles are smaller; Projects and History have ‹ › (⌘[ ⌘]) and how many; Projects switches list and cards in its header (⌥1 ⌥2). Toasts show at the foot of the page, except on the Studio, Dub, Stories and Audiobook, whose foot has a bar of its own: there they show in the header.

Updates show in the window

• Settings → About → Software updates: Check for Updates…, the daily check, installing without asking (a version found by the daily check installs and waits for the restart) and beta updates as before, and Show Changelog…, every version's notes in AudioSmith's language. The choices made in the old update window carry over.
• Releases send their notes in English and Chinese (CHANGELOG.md and CHANGELOG.zh-Hans.md), so actnow.dev's changelog page and the app show each in its language.

Discover voices: about 35 chosen voices instead of 1,184

• Discover voices now lists a small set chosen by listening tests' measurements (thread #91's audit): Chinese and English voices first, a few in Japanese, Korean, Spanish, French, German, Italian, Portuguese and Russian. The 1,126 generated combinations and the weaker hand-made voices are gone from the list.
• Each designed voice has a written character (age, voice, situation) that the model really gets, a seed that was tried, and a reference that ships with the app: Use voice saves it ready to speak, with no reference to wait for. Preview plays a clip made from that reference, as the Studio would say it; nothing is rendered to listen.
• The gallery also offers the built-in speakers worth starting from (Serena and Dylan for Chinese, Kokoro's best English and European voices): Use voice saves the speaker and opens the Studio.
• Filters: use (narration and audiobooks, conversation and podcasts, characters and kids, ads and announcing), language, gender and age, and search. Pitch, accent or dialect, style and the nine categories are gone.
• Voices saved from the earlier gallery keep working (they carry their own recipe), and dub casts or agents that name an earlier gallery voice or design preset still get it: those ids are kept, just not listed.

Design: eight starting points

• Narrator, Conversational, Newsreader, Presenter, Storyteller, Elder, Villain and the new Soft (which replaces Whisper and Calm Guide). Upbeat, Kid and Northeast Buddy are gone (a Northeastern voice is in Discover voices). None asks for slow speech any more.

Fewer weak built-in speakers

• Kokoro's af_nicole (a whisper) and bm_lewis (rough) are no longer offered, and Qwen3-TTS's Ryan is not offered in Chinese (he gave up on Chinese numbers in tests). A project or a setting that uses one keeps it, and the API still takes them by name.

Settings breathe, and say it shorter

• Every Settings page follows one spacing: the groups' padding, room above and below each row with a line between rows, a group's note as a block of its own, and room between titles, lines and controls. Writing help's model picker no longer cuts its value, and the status line that repeated the rows is gone.
• 28 texts are shorter, in English and Chinese.

Designed voices no longer say leftover reference text

• Slow designed voices (Storyteller, Elder, Meditation guide, and a few gallery voices) were cut off half-way through their reference: Qwen3-TTS stopped at a length limit that slow speech reached (engine patch 0008 raises it from 0.48 s to 0.8 s per text token, and AudioSmith's own limit can now raise it as well as lower it). The reference was still saved with its whole text, so clones said the missing half before every line.
• A reference made from a voice must now say its text: when the recognizer hears none of the three takes say it (or they were cut off), they are made once more; a kept take that says something else gets what it says as its text, or none is kept and the voice says why. Its text is never one the audio doesn't say.
• References made before are heard once, the first time a take, story or book clones them: one that doesn't say its text is made again (a note says so) or gets the text it says.
• Design prompts no longer ask for slow speech ("speaks slowly", "语速偏慢") from a preset or a picked pace: it made references slow and clones sound read. A description you write keeps it.
• Voices designed for Japanese, Korean, German, French, Spanish, Italian, Portuguese or Russian say their reference in that language, not English.
• Gallery previews use the tuned sampling the Studio uses (temperature 0.7, top-p 0.8 for Qwen3-TTS) instead of the engine's defaults; previews are made again once.
• The Hindi gallery voice is hidden: no model here designs or clones Hindi.

Dictation's sounds to choose

• Settings → Dictation → Sounds: the sound of listening starting, of it stopping and of an error are each chosen from the ActNow family's sound library (Tally, Wrap and Shrug by default), macOS's system sounds or none. A compact picker searches them (names and tags), filters by type, plays each with ▶ (again to stop) and sets the volume in steps; Try plays the choice.
• The sounds are React Bits Sound Lab's and come with the builds from actnow.dev and GitHub releases only: they are not in this repository or under its licence (NOTICE). A build from source plays the system sounds AudioSmith used before (Tink, Pop, Basso).

Models move to the ActNow folder, out of Time Machine

• Models are kept in /Library/Application Support/ActNow/models, shared with the other ActNow apps, instead of the Hugging Face cache (/.cache/huggingface), which cleaning tools treat as a cache. Models already there are brought over at the first start without downloading them again (hard links: the cache stays as it is, for the tools that use it).
• Time Machine leaves models out (they can be downloaded again); Settings → Advanced → "Back up models with Time Machine" turns it back on for all ActNow apps.
• A model file cut short (a disk error, a copy that stopped) is no longer taken for a downloaded model; downloading again repairs it. A download of another version left half-way no longer makes a downloaded model look incomplete. Checking a model's files waits for other ActNow apps working on it and tells them when a file was bad.
• If the ActNow folder's model claims are deleted (a cleaning tool), AudioSmith claims its models again when it comes to the front, and keeps models another app may still use.

Audiobooks keep their rendered parts

• The parts of an audiobook already rendered are kept in AudioSmith's data folder instead of ~/Library/Caches, which cleaning tools empty: Resume after a stop or a failure no longer renders hours of a book again. Those of earlier renders move over at the first start.
• Temporary files of recordings and imports cut short are cleaned at start.

Keys stay out of logs

• Google Translate (Basic)'s key is sent in a header, not in the address, so a network error no longer writes it into the log. Logs and the diagnostic bundle leave out keys, tokens and proxy passwords.

Dub editor: ⌘Z and ⇧⌘Z undo and redo again

• Menu items with ⇧ in their keys show and take exactly those keys (⇧⌘Z is Redo, not also ⌘Z).

The local writing model answers in full

• The built-in writing model (Qwen3 4B) could cut its answers short or garble them when they had to follow a fixed form. The AI director left the last line of a story undirected; the emotion director could lose half a scene ("surprisedn”: 2note”: in place of an emotion); Cinematic translation's review copied the translation instead of pointing out problems, so nothing was improved. These came from how the engine kept answers to the form. They are fixed: in tests every line of every story is directed and the review finds real issues.
• The same request now gets the same answer every time; before, the order of an answer's parts changed each time AudioSmith started.
• An answer that starts repeating itself is stopped at once and asked again, instead of running for minutes (a glossary note did).
• A retry after an unusable answer is no longer a copy of the same answer: it is sampled.

The apps share one copy of each model, some settings, and, if you agree, API keys, and they don't get in each other's way. AudioSmith names another app only when a released version of it is installed; with no other ActNow app, nothing on screen changes.

Downloads and updates move to actnow.dev

• AudioSmith now has its home on our website: actnow.dev/audiosmith. Download it from actnow.dev/download/audiosmith (free; the download needs a free account), and read every version's notes at actnow.dev/audiosmith/changelog. The source code stays on GitHub under the AGPL-3.0, and you can still build AudioSmith yourself.
• Updates come from actnow.dev: the daily check reads the version file there, and with Receive beta updates on also the betas' one, and offers the newer version. Copies of 0.4.0-beta.4 and earlier update to this version from GitHub, one last time.
• Settings → Diagnostics and about links to the website, the download page and the release notes.

Models shared with other ActNow apps

• On first run, if another app (say Dicta) has already downloaded some of the models AudioSmith needs, the setup says so ("3 models are already on this Mac · Installed by Dicta · no download needed") and uses them. Only the rest is downloaded. Choose models… lets you pick them yourself on the Models page.
• The Models page shows which models AudioSmith shares with other apps ("Shared with Dicta"), which ones another app has loaded right now ("Also loaded in Dicta"), and which ones another app installed ("Installed · from Dicta", with Use).
• Deleting a model that another app also uses asks what to do. Remove from AudioSmith keeps the files on this Mac for that app. Delete for all apps deletes them; the other app then has to download the model again. Clean up and "Erase everything" keep what other apps use, and say which app uses it. A model another app deleted for all apps shows "Deleted in Dicta · Download again".
• Settings → Storage shows how much space AudioSmith shares with other apps. It also offers to clean up models that only uninstalled apps used.
• If you accepted a model's licence in another app, AudioSmith still shows the licence, says when you accepted it, and asks you to continue.
• Kokoro's, Chatterbox's and Whisper's helper files now always load from the copy AudioSmith downloaded, at that version. Before, the engine could download them again on first use, or pick up another app's version.
• Downloads coordinate with the other apps: a file another app is downloading is not fetched twice ("Downloading in Dicta…"), and a model is not deleted while another app downloads or uses it.
• The engine no longer guesses which version of a model to load when several are on disk.

Settings shared with other ActNow apps

• The interface language, proxy, download source and model folder are shared with the other ActNow apps. With another app installed, each of these settings says whether it applies to all ActNow apps, and you can keep a different one for AudioSmith only.
• When another app keeps its models in a different folder, the Models page offers to move AudioSmith's models there. The move can be cancelled, and models the folder already has are not copied again.

API keys shared with other ActNow apps

• API keys and the Hugging Face token can be shared with the other ActNow apps. With another app installed, a checkbox under the key ("Let other ActNow apps on this Mac use this key") chooses whether to share it.
• AudioSmith offers a key saved in another app ("Use the DeepSeek key from Auralink"). It reads that key only after you click Use this key.
• Removing a key that another app also uses asks whether to stop using it here or to delete it for all apps. If another app deletes a shared key, AudioSmith asks you for a key again. "Erase everything" never deletes another app's key.
• The Hugging Face token is looked up in this order: $HF_TOKEN, the shared token, AudioSmith's own, the huggingface-cli token file.

Shortcuts and memory

• If another ActNow app uses the same keys as the dictation shortcut, Settings → Dictation says so. The app that took the keys last shows the notice once, with Change shortcut, Open Auralink settings and Ignore. While you record a new shortcut, "Also used by Auralink (Transcribe)" warns you, but you can still save it.
• The ActNow apps share one memory budget. Before AudioSmith loads a model, it unloads its own idle models first. If that is not enough, it asks the other apps to unload models that have been idle for a minute. It does the same when they ask, and never unloads a model a job is using.
• Dictation's paste no longer overwrites the clipboard while another ActNow app is pasting.

Dub: conversations recorded together (multi-speaker dialogue)

• A new switch in the Dub page's Cast step, 「多人对话配音」 (Multi-speaker dialogue): the people of a conversation are recorded together by the dialogue model "Dialogue · Light" (SoulX-Podcast), each line hearing the one before it, so answers sound like answers. Each line is then fitted to its own time as before (Smart fit and the other timings) and placed where it was said, so the video's pauses and overlaps stay. Monologues, one person speaking alone and lines over 4 s apart stay line by line.
• On by default for new dubs into Chinese or English with two speakers or more; dubs made before keep their audio (off). Each speaker is cloned from about 12 s of their own lines, with what they say there, even from another language (an English or Japanese video dubbed into Chinese); a saved or designed voice from its own reference.
• A line the dialogue model does not say clearly is made again (after the lines before it, another seed), then line by line; such a line is checked too, and marked "Listen" when it is not clear either. Editing a line records it alone again after the lines before it; changing a voice records the conversation again. The line list shows each conversation's lines with a coloured bar and a 「对话 N」 chip; each line's ⋯ menu can record the whole conversation again or keep a line out of it.
• Greyed with the reason on Macs under 16 GB, with the 「逐句」 voice match, and for other languages (a Japanese track stays line by line; a dub into Chinese and Japanese uses it for Chinese only). Not downloaded yet: the switch offers the download (2.4 GB), and dubs are made line by line until it is there. The Generate step's model picker lists it too. Slower than line by line.
• The translation of such a track aims a little shorter: the dialogue model speaks about 1.01–1.12× the estimate (learnt from the lines it makes).
• Local API and MCP: create_dub, edit_dub and batch_dub take dialogue (auto or off) and dialogue_model (soulx), edit_dub a line's no_dialogue; get_dub returns dialogue, dialogue_blocked, dialogue_stretches and each line's dialogue, made_by (dialogue or single), fallback_reason and flag; run_dub with segments naming a conversation line records it again after the lines before it.

The first signed release is a beta (a GitHub prerelease): copies that receive beta updates get the next beta, and 0.4.0 once it is out.

Updates

• AudioSmith updates itself: once a day it reads the version file of its latest release on GitHub and, when a new version is out, shows its notes (this file's section of the version) and offers to install it, skip it or remind you later. Updates are signed; AudioSmith checks the signature before it installs one, and usually downloads only what changed. AudioSmith → Check for Updates…, the command palette and Settings → Diagnostics and about check at once; that page also turns the daily check and installing without asking on or off. Copies built on your own Mac don't update themselves.
• Releases are signed with a Developer ID and notarized by Apple, so they open without Gatekeeper's warning (docs/release.md).
• Beta updates: Settings → Diagnostics and about → Receive beta updates. Betas get new features sooner but may be less stable; turned off, AudioSmith only updates to releases. It is on in a beta copy (such as this one) and off in a release until you choose. Turning it off never goes back to an older version: a beta stays until a release newer than it is out. About shows a Beta badge next to a beta's version.

Transcription with speakers

• New model "Transcription with speakers" (区分说话人转写; MOSS-Transcribe-Diarize 0.9B, 8-bit, 1.3 GB, Apache-2.0): it transcribes and tells who says what in one pass. Download it from the Models page, Transcribe or the Dub page; nothing downloads by itself.
• Transcribe: a Speakers option in the toolbar. Transcripts list their speakers (S1, S2…) with their lines; click one to name it. Copy and TXT export put one paragraph per turn after the speaker's name, SRT cues start with it, VTT marks voices, JSON has each segment's speaker. Dictation and plain transcription stay on Qwen3-ASR (MOSS is about 4× slower: a 10-minute recording takes about 4 minutes on a MacBook Air M4).
• Dubbing: with the model downloaded, a source with more than one speaker is transcribed by it and every line gets its speaker from it (in our tests 1.00 of the lines right on three of four 2–4 speaker dialogues and 0.93 on a 13-minute one with five people, where Qwen3-ASR + Sortformer managed 0.50–0.69 and never found more than 2 speakers; no limit of 4 speakers); the aligner still times the words, and the glossary's terms are its vocabulary. Without it, the dub works as before (Qwen3-ASR + Sortformer) and suggests the download, then Transcribe again.
• Long recordings go in 10-minute windows that overlap by a minute, linked by who speaks in the overlap; stretches without speech are skipped, and the model's chat answers to silence ("I'm sorry, I can't assist…") are dropped.
• Local API and MCP: speakers=true, model=gpt-4o-transcribe-diarize or response_format=diarized_json; segments carry speaker. MCP transcribe takes speakers: true.

Fewer transcription and writing models; Qwen3-ASR transcribes everything

• Speech recognition is Qwen3-ASR 0.6B ("Chinese & English transcription") everywhere: dictation, Transcribe's Fast and Accurate modes, dubbing and its read-back check, workflows, Convert, clone references and calls. It made 0.5–4 % errors on Chinese where Whisper made 8–33 %, punctuates, and is 2–4× faster (thread #44's measurements). Whisper large-v3 turbo ("Multilingual transcription", tagged Fallback) is used on its own for the languages Qwen3-ASR does not know (it knows 30) and for translating speech into English. Fast and Accurate now differ only in the word timings.
• Long files: the vocabulary is only sent with recordings up to 30 s (Qwen3-ASR copied it into the end of longer transcripts); the " The." (or "嗯。") Qwen3-ASR sometimes invents after the last sentence of a long file is dropped; files longer than about 4½ minutes are timed word by word in windows of at most 4 minutes, cut between sentences (the aligner's times broke down after about 4½ minutes).
• Removed: Parakeet TDT 0.6B v3 (no Chinese) → Qwen3-ASR, or Whisper for a language Qwen3-ASR does not know; Mel-RoFormer (its output ended in 2–3.5 s of silence, and it was 5–9× slower) → HTDemucs; Gemma 4 E4B, Gemma 4 E2B and Qwen3 1.7B (slower or broken JSON, weaker translations and clean-up) → Qwen3 4B Instruct 2507, now the only language model, named in Settings → Writing help instead of a choice. Settings that named a removed model go back to the default once, with a notice; a model you chose that is still there (Whisper) stays. API and MCP: a transcription that names Parakeet is answered by its replacement with replaced_by and a warning (X-AudioSmith-Replaced-By / X-AudioSmith-Warning); GET /api/models/{id} describes the removed models, and they are not listed. Their files stay until Models → Clean up, which now also lists them (and shows on their tab).
• Onboarding: Balanced no longer downloads Whisper; Max adds it instead of Mel-RoFormer.

Models named for what they do

• Every model has a plain name that says what it is for — Standard voices, High-fidelity clone, Fast clone, Voice design, Fast English voices, English clone, Multilingual all-rounder; Chinese & English transcription, Writing assistant · Standard… (标准音色、高保真克隆、快速克隆…) — a one-line description and a few tags (Chinese, Cloning, Built-in voices, Low memory, Non-commercial…). Pickers, the Models page, ⌘E, Settings, the bar beside Generate, History, onboarding, notifications and messages use them.
• The family, version, size, quantisation, repository and licence are still there: in a model card's Details on the Models page, in the pickers' tooltip and ⓘ, and in History's details (Technical name). Ids, repos and saved projects don't change.
• The model picker and the Models page have a search field that finds a model by its name, its technical name, id, repo or a tag.
• Local API and MCP: model lists add display_name (English), ui_name (as the app shows it), description and tags; a friendly name picks a model too.

Portable voices, and simpler Design, Clone and Export

• Make a portable copy of a model's built-in speaker or of a designed voice (My voices → ⋯, the Studio's voice list, or Make portable where a voice can't be used): its model says a ten-second line, evened out, and the take is saved as a cloned voice "<name> (approximate)" that every model that clones can speak with. It sounds close to the original, not the same, and says so.
• When a voice can't be used with a model, the message says why and what to do (make a portable copy, or pick a model that designs).
• Design shows only what the model can't express; what it will hear is behind ⓘ, and the seed moved into Details. Clone: style and language are under Advanced. Export audio: format, loudness, file name and extras in view; bit depth or bitrate, sample rate, channels and trimming under Advanced.

Simpler Stories, Audiobook and Dub

• Stories and Audiobook: the inspector keeps Setup, Cast, Performance (Stories) or Chapters and Cover & details (Audiobook), and one Advanced section for the rest — language, merging and evening out voices, refinement and guidance, gaps, trimming, seed, a stem per character, a story's M4B with its cover and details, an audiobook's own pronunciations — with N changes and Restore recommended (your pacing stays).
• The Project section moved to a Project ▾ menu in the header: Save (under the title — the title is the project's name), Save a copy, New, recent projects, All projects…
• A story shows only what its model hears: personas, the scene, the AI director and line directions disappear on models that say lines as written, and a line's speed on models without one.
• Dub: each step keeps its main choices; the rest folds into Advanced at the foot of the step (speakers and review, more languages and the translator's options, voice match, the voice model's parameters, a WAV's bit depth). The buttons the editor's ⋯ menu already had left the panels; Transcribe again… and Translate everything again joined that menu.

A simpler Studio

• The inspector shows what most takes need: the model, one Voice list (the model's speakers, your saved voices, a clip to clone, Manage voices…) and the speed. Everything else — style, reference transcript, length, refinement, guidance, sampling, seed, reading numbers, evening out the volume — is under Advanced, filled with the model's recommended settings; its header says "Recommended settings" or how many you changed, with Restore recommended inside. Advanced stays open or closed as you left it.
• The Voice controls and Audio quality buttons are gone, and so are the WAV precision (takes are kept at 24-bit; choose the file's bit depth in Export audio), Preload / Unload (in the ⌘E engine switcher, now with Unload), the autoplay switch (Settings → Studio) and the steps slider (three tiers remain).
• A clip chosen to clone is transcribed for you when it is a WAV.
• The bar beside Generate says which model and voice will speak, and how many settings you changed; a click shows the inspector.

Recommended models and settings

• Every speech model was measured on a MacBook Air (intelligibility with two recognisers, naturalness, voice stability, speed, memory). Each page now recommends two or three models for its job and your language, with a one-line "best for" and a speed and quality hint; the others are one click away under "Show all models", and the ones that can't do the job (no cloning for a dub) are greyed out with the reason.
• Tuned defaults: Qwen3-TTS uses temperature 0.7 and top-p 0.8 unless you (or an API request) set them — the models' own 0.9 / 1.0 sounded worse and less stable; OmniVoice uses guidance 1.5 with 32 steps (16 for dubbing). The Local API and MCP tools follow the same defaults. Rendered chunks from before are made again once with the new settings.
• Removed: six speech models that misread, paused or covered too little in the tests — KittenTTS mini, MOSS-TTS-Nano, Chatterbox (multilingual), Marvis, Soprano and Pocket TTS. Seven remain: Qwen3-TTS 0.6B CustomVoice, 0.6B Base, 1.7B Base and 1.7B VoiceDesign, Kokoro, Chatterbox Turbo and OmniVoice (under Show all models). Drafts, projects, dubs, workflows, settings, history and API or MCP calls that name a removed model use its replacement — Kokoro, Qwen3-TTS 0.6B Base for MOSS-TTS-Nano, Chatterbox Turbo for Chatterbox and for cloning with Marvis, Soprano or Pocket TTS — and the app says so once; the API answers with replaced_by and a warning. A removed model's own speakers saved in My voices can't be used any more.
• Models → "Removed models still take X GB": Clean up deletes their files (after listing them); nothing is deleted on its own.
• Chatterbox's expressiveness control is gone with the multilingual model (Turbo does not take it).
• New dubs fit lines with "Smart fit" instead of "Lip sync": lip sync sped most Chinese lines up audibly (1.2–1.4×). Dubs you already made keep their timing.
• Kokoro's six Chinese voices are hidden (unintelligible in the tests); its English voices are the best English default.

Stories: dialogue that flows

• Turn-aware pacing: the silence between lines follows the conversation — shorter when a character goes on or answers a question, a little longer after a statement, shortest when a line is cut off (… —), longer around narration. Three presets (Dialogue, Reading, Relaxed) in the new Performance section; new stories use Dialogue, stories made before keep their fixed gap until you pick a preset. A line's own "pause after" still wins.
• Consecutive lines of one character are said in one go, so the voice flows from one sentence into the next, then split back into lines at their pauses (player, stems and the sidecar see lines as before; a turn that cannot be split is made line by line). The pause inside a turn is kept close to the preset's.
• Each character's lines are brought toward one loudness (measured 4–7 LU apart on Qwen3-TTS before, about 1 after).
• A scene for the story, a persona for each character and a direction (with an intensity) for each line. Qwen3-TTS CustomVoice and VoiceDesign hear them as an instruction with every line; other models say lines as written, and the page says so ("Line by line"). Never spoken.
• AI director: the writing-help language model (Qwen3 4B on this Mac by default) reads the story and fills in the scene, personas and directions; edit any of them, Undo, or direct again.
• Imported scripts keep stage directions out of the spoken text: 妈妈（温柔地）：…, MOM (softly): …, [Mia] (pleading) … become the line's direction.
• Agents: update_longform sets a story's scene, personas, directions, intensities and pacing, and direct: fill | all runs the AI director; get_longform shows them with the model's continuity level.
• The catalogue marks Qwen3-TTS CustomVoice and VoiceDesign as taking instructions, as the engine reports them; dubs and stories ask one predicate.

Export audio

• History's takes have an Export button (and Export… in ⋯); the Studio's latest take, a take's details and Projects open the same Export audio dialog, which replaces the Save as… menu.
• Choose the format (WAV; M4A through the engine; MP3, FLAC and Opus with ffmpeg, greyed out with a pointer to Settings → Audio tools without it), the bit depth or bitrate, the sample rate and the channels; bring the file to −16, −18 or −23 LUFS with a true-peak ceiling; trim the silence at both ends; name the file and its title and artist tags; save the script and subtitles beside it. The dialog shows the size and length before you export, and remembers your choices for each kind of take (Studio, voice preview, converted recording, Local API).

Writing help no longer fails during a render

• The language model on this Mac now waits for its turn when a render, dub or batch is using the engine, instead of failing with "the engine is busy": writing help, the AI director, dub translation, transcript clean-up, the direction coach, phone calls and API/MCP requests go next, ahead of queued batch work (a running job is never interrupted). These short requests don't appear in the jobs list. Their time limits count from when the model starts.
• Dictation never waits for a render: while one runs, it types the polished text at once without the model's clean-up. The Tools direction coach shows 「等待当前任务完成…」 (Waiting for the current job to finish…) and, if the job is still running after 30 s, reads the note by keywords.