* Fix voice sample validation on Python 3.13
Python 3.13 removed audioop from the standard library, which broke reference
audio validation when adding voice samples. Add the audioop-lts backport for
3.13+ installs and bundle audioop in PyInstaller builds on the same versions.
* style(tests): satisfy Ruff import ordering
---------
Co-authored-by: Jamie Pine <[email protected]>
* feat(windows): add native ROCm support for AMD GPUs
Implements native ROCm architecture for Windows.
- Adds backend build pipeline for voicebox-server-rocm.exe
- Detects AMD GPUs dynamically and routes PyTorch allocations
- Adds automatic download and update logic for ROCm dependencies
- Refactors UI in GpuPage.tsx and GpuAcceleration.tsx to add AMD flows
- Fixes 'Switch to CPU' lock on Windows via Tauri backend_override state
- Resolves PyInstaller/rocm_sdk UnboundLocalError silent crashes
- Resolves Numba/NumPy 2.x incompatibilities during Qwen3-TTS load
- Resolves HF_HUB_OFFLINE Catch-22 for CustomVoice processor caching
* fix(rocm): host libs archive under the app release tag, drop offline-load regression
Align the ROCm libs download with the CUDA pattern: both the server core and
the libs archive are published under the app-version release tag, with the libs
content version encoded in the filename only. The previous code fetched libs
from a separate rocm7.2-v1 tag, which disagreed with the download test.
Also revert the unrelated Qwen CustomVoice changes that wrapped model loading in
force_offline_if_cached (not imported — a NameError on load for every platform)
and re-added a Base-model cache gate. The inference-path offline guard was
deliberately removed previously.
* feat(rocm): gate download on AMD detection and persist the backend variant
The ROCm download section now only shows when the backend reports an AMD GPU on
Windows (new supports_rocm health field, backed by the memoized
is_amd_gpu_windows detection that was previously unused), or when ROCm is already
downloaded/active.
Make the backend override honor a pinned variant: set_backend_override persists
the choice to disk so it survives an app restart, start_server reads it back,
and a cuda/rocm pin now actually selects that variant instead of always
preferring ROCm. A stale pin to a deleted backend self-heals to the default
order rather than forcing CPU. Add the web no-op stub for the new method.
* chore(rocm): drop incomplete vitest harness for the unused GpuAcceleration component
GpuAcceleration.tsx is not routed anywhere (GpuPage is the live settings view),
and the added vitest setup referenced testing-library/vitest deps that were not
in the lockfile, breaking the web typecheck. Remove the dead component's test
and its scaffolding to keep this PR scoped to the ROCm feature.
* ci(rocm): add ROCm release-artifact pipeline
Mirror the CUDA packaging path for ROCm so the runtime download has artifacts to
fetch. scripts/package_rocm.py splits the PyInstaller --rocm onedir into
voicebox-server-rocm.tar.gz (core) + rocm-libs-rocm7.2-v1.tar.gz (AMD runtime:
HIP DLLs, rocBLAS Tensile data, MIOpen kernel DBs) + rocm-libs.json, matching
the names services/rocm.py expects, both under the app-version release tag.
The new build-rocm-windows job in release.yml builds on windows-latest/cp312 and
lets build_binary.py --rocm pull the official AMD Radeon wheels.
The file classifier can't be validated against a real AMD build on CI, so it has
unit coverage (test_package_rocm.py) against a synthetic onedir layout. The
prefixes/dir markers may need a tweak after the first real build on AMD
hardware — the packager hard-fails loudly if it classifies zero ROCm files.
---------
Co-authored-by: Jamie Pine <[email protected]>
* feat(capture): dictation, personalities, 0.5.0
Ships the Capture release end to end. Global-hotkey dictation with
synthetic paste into the focused app on macOS and Windows, an on-screen
pill across recording / transcribing / refining, customizable push-to-
talk and toggle chords, and an accessibility-permission prompt scoped to
Settings → Captures with inline re-check feedback.
Voice profiles gain optional personalities that power compose / rewrite /
respond actions via a local Qwen3 LLM — shared with refinement, so there
is one local LLM in the app, not two.
Refinement hardened with deterministic Whisper-loop collapse before the
LLM sees the transcript, per-capture flag snapshots for re-runs, and a
ten-transcript evaluation harness across every bundled refinement size.
Version bump 0.4.5 → 0.5.0.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* feat(mcp): local MCP server exposes voicebox.* tools to AI agents
Mounts FastMCP at /mcp (Streamable HTTP) so Claude Code, Cursor,
Windsurf, and the VS Code MCP extensions can call voicebox.speak,
voicebox.transcribe, voicebox.list_captures, and voicebox.list_profiles
against the running Voicebox server.
Backend
- new backend/mcp_server package (tools, middleware, profile resolve,
pub/sub events); named mcp_server to avoid shadowing the installed mcp
PyPI package FastMCP imports internally
- app.py migrated from @app.on_event to lifespan= so FastMCP's session
manager cohabits with Voicebox's startup/shutdown
- new MCPClientBinding table + /mcp/bindings CRUD; ClientIdMiddleware
reads X-Voicebox-Client-Id into a ContextVar and stamps last_seen_at
- profile resolution precedence: explicit -> per-client binding ->
capture_settings.default_playback_voice_id
- POST /speak REST wrapper for non-MCP callers (shell, ACP, A2A)
- GET /events/speak SSE broadcasts speak-start / speak-end so the pill
surfaces agent-initiated speech
- backend/mcp_shim proxy (plain httpx) for stdio-only MCP clients
- PyInstaller spec updates + new --shim build target (~18 MB)
Frontend
- Settings -> MCP page with HTTP / stdio / claude-mcp-add copy snippets,
default voice picker, per-client bindings table, connection status
- useMCPBindings, useSpeakEvents hooks
- CapturePill gains 'speaking' state; DictateWindow subscribes to SSE
and emits dictate:show so the Rust side surfaces the pill window
Native
- tauri.conf.json externalBin now includes voicebox-mcp
- show_dictate_window helper + dictate:show listener in main.rs
- (also in this commit: InputMonitoringGate UX, hotkey_monitor tweaks,
landing footer/navbar updates, new overview docs for captures /
dictation / mcp-server / voice-personalities)
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* feat(mcp): Rust-owned speaking pill with self-contained audio playback
The pill window now surfaces for agent-initiated speech without main-window
involvement. Rust subscribes to /events/speak via a tokio task + reqwest
streaming body (speak_monitor.rs), shows the pill, and forwards events to
the dictate webview over Tauri's event bus. The pill plays audio via a
plain HTMLAudioElement and emits dictate:hide when playback ends. The
pill stays hidden through the ~1 s generation wait and only surfaces when
audio actually starts, with the counter armed at that moment.
Fixes a shared-dict mutation in mcp_server/events.publish() that caused
the second subscriber (Rust speak_monitor) to receive `event: message`
instead of named speak-start/speak-end frames. Also teaches the speak_monitor
parser to handle CRLF framing (sse-starlette default). Main-window
AudioPlayer now skips autoplay for source in {mcp, rest} to avoid
double-play when both windows are alive.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* readme and dev script
* feat(capture): gate global hotkey on dictation readiness checklist
Stops the "stuck pill" failure where pressing the chord with missing
STT/LLM models triggers a recording that has nowhere to land. The
hotkey now stays disarmed until every gate (models downloaded, Input
Monitoring + Accessibility granted) is green; the empty-state checklist
in CapturesTab surfaces each unmet gate with a one-click action and
auto-arms the chord once everything turns green.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* color
* model download status
* progress
* personality: bool API, i18n across the app
- Collapse intent tri-state (respond/rewrite/compose) to `personality: bool` on /generate, /speak, and voicebox.speak. Drop respond entirely; keep compose as a standalone button via /profiles/{id}/compose. Remove /rewrite, /respond, and /speak profile endpoints.
- FloatingGenerateBox: Wand2 persona toggle + Dices compose button appear when the selected profile has a personality. ProfileCard badges Wand2 alongside the effects Sparkles.
- MCP bindings: default_intent column → default_personality: bool. Migration drops the legacy column.
- i18n: en / ja / zh-CN / zh-TW translation files filled out and wired through the capture, server, and profile UI.
```ts
voicebox.speak({
text: "Deploy complete.",
profile: "Morgan",
personality: true, // rewrite through the profile's personality LLM
});
```
* i18n: GenerationPage sidebar copy
* fix: BOOL import for windows crate 0.62
BOOL moved from Win32::Foundation to windows::core in 0.62.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
* fix(capture): layout-aware V keycode for synthetic paste on macOS
macOS apps match Cmd+V against the layout-translated character via NSMenu
key equivalents, so posting kVK_ANSI_V (= 9, the QWERTY V position) on
Dvorak produces Cmd+. and never triggers Paste. New keyboard_layout
module resolves the active layout's V keycode via
TISCopyCurrentKeyboardLayoutInputSource + UCKeyTranslate, caches it in an
AtomicU16, and refreshes on kTISNotifySelectedKeyboardInputSourceChanged.
All TIS calls run on the main thread (init from Tauri setup; observer
callback delivered to the main runloop); synthetic_keys::send_paste
reads the cached value once per paste. Falls back to kVK_ANSI_V when
resolution fails or the active input source carries no Unicode key
layout data.
Windows is intentionally left on hardcoded VK_V — SendInput delivers
WM_KEYDOWN with wParam = VK_V to the target regardless of the active
layout, which is why `Send "^v"` works for AutoHotkey on Dvorak Windows.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(capture): cooperative app activation for synthetic paste on macOS 14+
macOS 14 deprecated NSRunningApplication.activateWithOptions: in favour
of a cooperative-activation pattern: the caller first yields activation
rights to the target, then the target activate()s against the tightened
Sonoma foreground rules. Without the yield, activate() on 14+ sometimes
silently fails or only bounces the dock icon — the exact "paste lands in
the wrong app" symptom we were previously one API break away from.
activate_pid now discovers the 14+ selector via respondsToSelector: and
branches: on 14+ it yieldActivationToApplication:'s from
NSRunningApplication.current then calls -activate on the target; on
11–13 it stays on -activateWithOptions: (still the only option). Both
branches propagate the BOOL return — if activation is refused we error
out before clobbering the clipboard instead of silently proceeding.
The respondsToSelector: result is cached in a OnceLock so the probe
isn't repeated on every paste.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(capture): conditional clipboard restore + always-attempt on paste failure
Two bugs in paste_final_text' clipboard handling:
1. Restore was unconditional. If the user ⌘C'd in the target app during
the 400 ms paste-consume window — or a clipboard history tool (Paste,
Pastebot, Maccy) or Universal Clipboard sync snapshotted our staged
text — the blind restore overwrote their newer content with the
pre-paste snapshot, silently losing user data.
2. send_paste' errors were propagated with ? before the restore, so a
CGEventPost / SendInput failure left the user's clipboard stuck on
the transcript.
Fix folds both into one pattern: capture the post-write change count,
re-read it after paste-consume, restore only when they match (plus treat
a change-count read failure as "unknown, don't overwrite"). Isolate
send_paste's error so the restore runs regardless of paste success, then
propagate the paste error after.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* chore(deps): pin rdev to jamiepine/rdev fork
Upstream Narsil/rdev has shipped no release since 2023-06 (crates.io
still serves 0.5.3), so the Sonoma main-thread fix we depend on — PR
#147, applied at hotkey_monitor.rs:184 — is only reachable via a git
pin. A pin to a third-party repo breaks the build whenever the remote
force-pushes, renames, or is taken down, and Cargo does not durably
cache git-dep archives the way it does crates.io tarballs.
Forking to jamiepine/rdev at the same SHA removes that failure mode
without changing crate behavior and gives us a place to cherry-pick
future OS-compatibility fixes on our own timeline. The SHA was verified
to exist on the fork before re-pinning.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(mcp): idle timeout + escalating backoff for speak-SSE monitor
Two reliability gaps in the /events/speak subscriber:
1. resp.chunk().await had no idle timeout. A backend that accepts the
TCP connection but stops producing frames (deadlocked SSE endpoint,
zombie process) would block the task forever without reconnecting.
The pill window would never surface for agent-initiated speech and
there would be nothing to log. Backend emits a `:ping` heartbeat
every 15 s, so 45 s without any data is now treated as a dead
stream — the task errors out and the reconnect loop takes over.
2. Flat 2 s backoff escalates nowhere. Logs fill with reconnect lines
when the backend is down for minutes, and a backend that accepts +
immediately closes connections (no data) spins the loop tightly.
Backoff now escalates 500 ms → 30 s on unproductive rounds and
resets only when at least one frame arrives (the connection was
genuinely productive, not just accepted).
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(refinement): character-level loop collapse + pytest coverage
The word-level pass catches single-word Whisper loops ("URL URL URL…")
but misses two common hallucination patterns the PR had to claim as
"edge cases":
1. Multi-word English loops — "thanks for watching thanks for watching…"
× 6 sails through because no two consecutive tokens are identical
after text.split().
2. CJK loops — "謝謝觀看" × 7 sails through because text.split() returns
a single unsplit token for the whole loop (no whitespace between
characters).
Add a character-level second pass: a non-greedy regex finds any 2–60
char substring that repeats min_run+ times immediately after itself and
strips the run. The 2-char floor keeps emphasised single-letter runs
("wooooooow") intact. The 60-char ceiling covers every observed
Whisper tail hallucination ("Please like and subscribe to my
channel.", "Subtitles by the Amara.org community") while staying short
enough that coincidental long-phrase repetition in legitimate speech
doesn't hit the threshold. Whitespace normalisation only runs when the
pass actually stripped something, so untouched transcripts keep their
original spacing.
New test_refinement_collapse.py gives the pre-processor its first
deterministic unit-test coverage: 17 tests pinning the word-level
legacy behaviour plus the new multi-word English / CJK / Japanese /
emphasis-preservation cases.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(db): graceful fallback when SQLite < 3.35 on MCP bindings migration
SQLite gained ALTER TABLE … DROP COLUMN in 3.35 (Mar 2021). Production
PyInstaller builds bundle Python 3.12 which links to SQLite 3.40+ so
that path is always safe, but a dev running the backend directly on
Ubuntu 20.04 (3.31) or Debian 11 (3.34) would crash on first startup
trying to drop the legacy default_intent column.
Add _supports_drop_column(engine) — returns True on non-SQLite
dialects (Postgres / MySQL have supported DROP COLUMN for decades) and
gates on the runtime sqlite_version for SQLite. When unsupported, log a
warning and leave the unused column in place: SQLAlchemy only maps
declared columns, so a stray default_intent column does no reads or
writes and can't interfere with runtime behaviour.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(mcp): correct lifespan shutdown order — drain MCP before unloading models
The inline lifespan ran _run_shutdown inside the MCP context, so the
TTS / Whisper / LLM models were unloaded *before* FastMCP's __aexit__
got a chance to cancel its in-flight session tasks. Any MCP request
mid-generate at shutdown time would crash on "model unloaded" instead
of receiving a clean session-cancelled error.
Rewire via compose_lifespan (which was already defined in
mcp_server.server for exactly this purpose but never used):
AsyncExitStack enters factories in order and exits in LIFO, so
MCP teardown fires first — cancelling sessions — and _run_shutdown
runs after nothing is holding the models. Smoke test shows the
log order flipped as expected:
Ready
StreamableHTTP session manager started
... running ...
StreamableHTTP session manager shutting down ← was last, now first
Voicebox server shutting down... ← was first, now last
As a side benefit, _run_shutdown is now paired with _run_startup via
try/finally inside voicebox_lifespan, so a partial startup (models
half-loaded, MCP __aenter__ fails) still unloads whatever was loaded
instead of leaking it to process exit.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(mcp): stamp last_seen_at on /speak too + tighten path predicate
POST /speak is a REST wrapper around voicebox.speak for agents that
don't talk MCP (shell scripts, ACP, A2A). It reads X-Voicebox-Client-Id
and uses it for the same per-client profile resolution + default
personality lookup the MCP tool does (speak.py:39-64), so its callers
are first-class clients — but the ClientIdMiddleware only stamped
last_seen_at on /mcp* paths. REST speak callers showed up as "never
seen" in Settings → MCP despite actively acting on their bindings.
Widen the stamp predicate to an explicit ("/mcp", "/speak") prefix
list, and require a path boundary on match so future routes named
/mcpfoo or /speakers don't silently inherit the stamp via the prefix.
New test_client_id_middleware.py pins the scope with 17 parametrised
cases (both the allowed set and the overlap cases that must not match).
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* feat(captures): scrubbable WaveSurfer player for capture detail view
Replace the placeholder fake-waveform + play button in CapturesTab's
audio card with a real CaptureInlinePlayer (wavesurfer.js). The player
renders the actual waveform, lets users scrub through the clip, and
shows a proper current/total timestamp pair in place of the
duration-only label.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* feat(ui): persist selectedProfileId across sessions
Wrap useUIStore in zustand/middleware's persist under the key
voicebox-ui. partialize only selectedProfileId so volatile UI state
(dialog open flags, form drafts, engine/voice pickers, sidebar) stays
in-memory as before — but reopening the app no longer loses whichever
profile the user was last working with.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* feat(captures): mirror readiness checklist into the settings sidebar
The six-gate checklist only rendered in the CapturesTab empty state, so
a user already on the settings page had no single surface showing which
gate was red — the inline InputMonitoringNotice covered one, the model
pickers covered another, and Accessibility was only hinted at by the
auto-paste toggle. Mirror the same component into the right sidebar of
the settings page so every gate (STT model, LLM model, Input Monitoring,
Accessibility, plus the hotkey toggle in the main column) is always
visible while the user configures dictation.
New compact prop on DictationReadinessChecklist drops the centered
header and empty-state max-width so it fits the 280 px sidebar next to
the existing About / Differences blocks. Callers in compact mode own
the heading — CapturesPage reuses the existing captures.readiness.title
key (present in en / ja / zh-CN / zh-TW already) as an h3 matching the
sibling sidebar sections.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* feat(captures): move sidebar checklist below differences + hide when all green
Two small follow-ups to the sidebar checklist placement. Move it below
the What's different section so the sticky top of the sidebar stays the
page's narrative context (About → differences) and the checklist reads
as a status panel rather than preamble. Gate the whole block on
!readiness.allReady so once every gate is green the sidebar drops back
to just About + What's different — no value in real estate full of
checkmarks.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(captures): refetch readiness immediately after STT/LLM model swap
useCaptureSettings updated its own cache optimistically but never
invalidated ['capture-readiness'], so for up to 5 s (the poll interval)
after switching stt_model or llm_model the checklist kept showing the
previous model's ready/missing state. The backend endpoint resolves
the model live on each call — it was just the frontend cache that
lagged. Invalidate in onSettled only when the patch touched a model
field, so unrelated updates (chord keys, toggles) don't pay for a
refetch.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(captures): hide macOS-only copy when running on Windows / Linux
Two surfaces leaked macOS-specific copy onto other platforms:
1. The Input Monitoring + Accessibility rows in the readiness
checklist rendered everywhere. On Windows/Linux the Rust permission
stubs return true, so the rows showed as permanent green checkmarks
with copy like "macOS allows Voicebox to detect your global
shortcut." — nonsense when you're on Windows. Gate both rows on a
userAgent-based isMacOS check so they only render where the
underlying TCC permission actually exists.
2. The global-shortcut setting description ended with "macOS will ask
for Input Monitoring permission the first time you turn this on."
That sentence rendered on every platform. The readiness checklist
already surfaces the TCC requirement at the right moment on macOS,
so the description doesn't need the platform note — drop it from
en / ja / zh-CN / zh-TW.
Other macOS strings (AccessibilityNotice, InputMonitoringNotice, their
"stillMissing" hints) are already gated behind the Rust permission
booleans returning false, which never happens on Windows/Linux, so they
stay inert without further changes.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* feat(capture): swap the rdev fork for keytap 0.2, delete local chord state machine
Dep swap:
- Drop the git-pinned jamiepine/rdev fork we were carrying since the
upstream crate is abandoned.
- Depend on keytap 0.2 from crates.io — our own cross-platform global
keyboard tap crate. Clean shutdown via Drop, Sonoma-safe by design
(no TSMGetInputSourceProperty calls off the main thread, so
`set_is_main_thread(false)` is gone), and properly versioned.
Chord engine rewrite:
- Delete hotkey_monitor.rs's internal Chord state machine (Match enum,
KeyEvent enum, step()/classify() methods, associated unit tests).
keytap's ChordMatcher subsumes it: Momentary chord for PTT,
add_toggle() for Toggle-to-talk, longest-match resolution, sticky-end
for Toggle. Net: -80 LOC in hotkey_monitor.rs; the remaining module
is the dispatcher loop + Effect→Tauri translation.
- Preserve the PTT→Toggle "RestartRecording" upgrade signal. keytap
emits End(PTT)+Start(Toggle) atomically (same Instant) when the held
set upgrades from a shorter chord to a longer superset. The
dispatcher peeks at the matcher with a 5 ms recv_timeout after any
End and coalesces the pair into Effect::RestartRecording so the
frontend still gets the "discard the transition-moment audio" signal
instead of an unrelated Stop+Start pair.
- HotkeyMonitor::update_bindings now actually tears down the tap on
empty bindings instead of leaving an idle CGEventTap around. New
bindings rebuild the matcher and the dispatcher thread from scratch.
key_codes.rs:
- Rewrite the browser-code → Key table against keytap's cleaner Key
variant names (`A`..`Z` not `KeyA`..`KeyZ`, `Digit0`..`Digit9` not
`Num0`..`Num9`, `ArrowUp` not `UpArrow`, `AltLeft`/`AltRight` instead
of `Alt`/`AltGr`, `Period` not `Dot`, …). On-disk chord string
format (W3C `KeyboardEvent.code` identifiers) is unchanged, so
capture_settings rows written before the swap round-trip identically.
Legacy aliases (`Alt`, `AltGr`, `Num0`, `UpArrow`, `Dot`, …) kept for
forward-compat on old rows.
main.rs / input_monitoring.rs:
- Update the few doc comments that referenced `rdev::listen` to
describe keytap's Tap; no behavioural change.
- build_chord_bindings now imports from keytap::Key.
- enable_hotkey / disable_hotkey / update_chord_bindings reach into
HotkeyMonitor via &mut since apply()/update_bindings() now mutate.
Tests live in keytap now (22 chord-related tests in keytap 0.2,
including the PTT→Toggle upgrade scenario that used to be tested in
hotkey_monitor.rs). Voicebox's hotkey_monitor.rs is thin enough that
local testing would be trivia.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* chore(deps): bump keytap 0.2 → 0.4 for macOS modifier-events fix
0.2 read CGEventFlags via CGEventGetIntegerValueField(event, 0x81),
which is not a valid CGEventField id — macOS silently returned 0, so
FlagsChanged events produced no KeyDown / KeyUp for any modifier key
and the PTT / toggle chords never armed on macOS. 0.4 uses the
documented CGEventGetFlags(event) API.
0.3 (tracing / serde / Fn / IntlBackslash) is picked up as a free
consequence; no API surface we depend on changed.
* perf(captures): stop polling readiness once both models are green
useQuery was firing GET /capture/readiness every 5s forever, and also
on every window focus. Once stt.ready and llm.ready are both true the
answer can only change when the user swaps a model in settings, and
useSettings already invalidates the query on that path — the polling
was pure noise.
Gate both refetchInterval and refetchOnWindowFocus on "not fully ready"
so we fall silent once the checklist is green.
* feat(ui): theme settings, stories polish, track editor restructure
- add dark/light/system theme with persisted choice + OS change listener
- restyle stories sidebar (search, item layout, border) to match captures
- move floating generate box to right column of stories, add top fade mask
- story track editor: sticky track labels aligned via flex rows, custom scrollbar with left/right zoom handles
- capture pill light mode pass, fix inline waveform progress color
- pull mcp_server hidden imports into the pyinstaller spec
- notarization doc draft
* fix mlx llm bundling
* feat(ui): shared ListPane primitive + misc polish
ListPane is a compound component (Header / TitleRow / Title / Actions /
Search / Scroll) that owns the relative wrapper, faded right divider
(50px top fade), top scroll mask, and absolute-positioned header used by
every list-detail tab. Wires up CapturesTab, StoryList, and EffectsList.
EffectsTab gets -mx-8 / pr-8 to match the edge-to-edge layout used
elsewhere.
Other changes:
- MCPPage: native <select> → shadcn <Select> for default voice and
per-binding voice pickers
- Button outline variant: add hover:border-accent
- Drop hover:text-destructive from trailing delete buttons
(HistoryTable, GpuAcceleration, GpuPage, EffectsChainEditor,
EffectsDetail)
- HistoryTable empty state moved behind t('history.empty')
- StoryContent scroll padding pt-14 → pt-16
- backend health reports the captures dir
- landing CapturesMockup: "Send to" → "Export" with Download icon
- CHANGELOG: drop [Unreleased] personality section
* fix(captures): Play As autoplay + default voice + orphan recovery
- Hand /generate ids to the global SSE watcher so playback fires on completion. The mutation onSuccess was checking audio_path on a queued row, which is always empty — autoplay never ran.
- Bind the Play As voice selection to capture_settings.default_playback_voice_id, kept in sync with the Settings → Captures and Settings → MCP pickers. Picking from the split-button dropdown writes back to settings.
- Extract AudioBars from HistoryTable into a shared component; use it for the Play As generating state in place of Loader2.
- Stop the active-state hover from flashing white text when the button is in its lighter accent/10 fill.
- Drop the gradient avatar swatches from the Settings → Captures voice dropdown.
- Backend: when the gen worker exits without writing a terminal status (e.g. SQLite lock racing the failed-status write inside its own exception handler), the cancel endpoint now flips the row to failed instead of 409-ing. Worker also force-fails on its way out as a belt-and-suspenders.
* fix(captures+chord): Stop button stops, ChordPicker accepts shorter chords
Two unrelated correctness bugs caught in PR review:
- The Play As "Stop" button was wired to handlePlayAs() unconditionally, so clicking it during playback kicked a fresh generation instead of halting. Now pauses the player when the click came from the main button while playbackState is 'playing'. Picking a different voice from the dropdown still kicks a new generation as before.
- ChordPicker tracked the peak set of held keys but seeded the peak from initialKeys, so a user who opened the picker with a 3-key chord saved couldn't replace it with a 2-key chord — the candidate length never beat the seed. The peak now resets on the first press of a fresh sequence (when no keys were held immediately prior), then grows monotonically within that hold.
* fix(settings): honor explicit null on nullable fields, ignore on the rest
Routes were calling model_dump(exclude_none=True), which drops every
client-sent null before it reaches the service. The service then layered
on its own `if value is not None` guard. Net effect: setting a nullable
column back to null was a no-op — the MCPPage default-voice picker sends
null when the user picks "no default" and the row was silently keeping
whatever was there before.
Switched the routes to exclude_unset=True so absent fields stay absent
but explicit nulls survive the dump, and centralised the per-field
nullability check in the service. The check inspects the SQLAlchemy
column metadata so non-nullable columns (stt_model, llm_model, the chord
key lists) still drop nulls instead of crashing the request, while
default_playback_voice_id can finally be cleared.
* fix(captures): clean up audio files when create_capture fails
The create flow wrote raw audio (and a transcoded .wav for non-wav
sources) to data/captures before the DB row was committed, so any
failure between the write and the commit — a webm that decoded to a
0-length array, a whisper model that errored mid-transcribe, a SQLite
contention on the commit — left the audio on disk with nothing pointing
at it. Over enough flaky uploads the directory grows without bound.
Now every path written before the commit is tracked in a list, and the
whole stretch from the first write to db.commit() runs inside a
try/except that unlinks each tracked file on raise and re-raises. The
transcode branch removes the raw file from the cleanup list only when
the unlink actually succeeds, so an OSError on the raw-path delete
still hands cleanup the original blob to retry.
* fix(mcp): restrict voicebox.transcribe(audio_path=...) to loopback
audio_path mode took any absolute filesystem path and returned its
decoded contents as transcribed text with no caller verification beyond
the existence/size checks. The X-Voicebox-Client-Id middleware records
the header but never rejects an absent or fake one, so a Voicebox bound
to 0.0.0.0 (the documented "remote access" mode) was effectively an
unauthenticated arbitrary-local-file read primitive.
The middleware now stashes the request's remote address in a ContextVar
alongside the existing client_id, and audio_path mode refuses anything
that doesn't parse as a loopback address (IPv4 127.0.0.0/8, IPv6 ::1).
audio_base64 mode is unchanged — that path was always bounded to bytes
the caller already has.
Loopback callers (the Tauri webview, local CLI scripts, MCP clients on
the same machine) keep working. Remote callers now have to send the
audio over the wire if they want it transcribed.
* fix: PR review nits — response shape, landing copy, form reset
- /llm/generate's "model is downloading" branch was raising HTTPException(202, detail={...}), which wraps the payload in {"detail": ...} and forces clients to parse a success status as if it were an error. Switched to JSONResponse so the payload sits at the top level.
- The landing page's "Language Models" card advertised "Qwen 3.5" with sizes 4B/2B/0.8B; we ship Qwen3 at 0.6B/1.7B/4B. Aligned to what's actually in the binary.
- ProfileForm's discard-draft button reset the form without touching `personality` or `avatarFile`, so stale persona text and an attached avatar would survive the discard. The other three resets in the file already include both fields — this brings the discard path in line.
* perf(mcp): move last_seen_at stamp off the request path
ClientIdMiddleware was running the SQLAlchemy SELECT/INSERT/UPDATE/COMMIT
inline on the event loop after every /mcp/* and /speak request. SQLite
serialises writes, so concurrent MCP traffic queued behind the stamp
write — the response sat waiting on a side-effect that the client never
needs in band, and SSE streams would stall briefly per request.
The middleware now hands the stamp to asyncio.to_thread via a fire-and-
forget create_task so the response returns immediately and the write
runs on the default executor. A module-level set keeps strong refs to
in-flight tasks (per asyncio docs) so the GC can't collect them mid-
write. The fallback path runs the stamp inline if no loop is available
(tests/oddball callers) rather than silently dropping it.
* fix(dictate): force-dismiss the speaking pill when SSE never comes back
The pill subscribed to /generation/{id}/status to know when to start
playback, but EventSource.onerror was a no-op — auto-reconnect was the
intended recovery for transient drops. The gap: if the backend deletes
the gen row mid-flight or the connection silently dies in a way the
browser keeps retrying without ever getting a status event, the pill
sits in 'speaking' forever and the user has no way to clear it.
Added a 60-second hard cap that arms when the SSE opens and clears the
moment any real status event lands. If it fires while the pill is still
on the same id and audio never started, it force-dismisses. Same idea
as the existing post-speak-end 15s grace, but covers the case where the
backend never says anything at all.
* fix: i18n cleanup + readiness checklist effect cadence + ChordPicker shadow
- DictationReadinessChecklist was constructing downloadByModel as a fresh Map every render and listing it in the cleanup effect's deps. With the 1 s polling cadence and arbitrary parent rerenders the effect ran more often than it needed to. Memoised the Map on activeTasks; the effect now keys off the memo's identity.
- zh-CN persona tooltipActive/ariaLabelActive matched their inactive twins byte-for-byte ("以人物设定朗读"). The other locales differentiate the active state with a -ing / -中 suffix; zh-CN now reads "正以人物设定朗读" when active.
- personalityPlaceholder was a ~290-character paragraph that doubled as both the example text and the explanation, repeating most of what personalityHint already said. Trimmed to the example only and folded the explanation + leave-blank consequence into the hint, across all four locales.
- Refinement model size keys were size06 / size17 / size4. Renamed the 4B variant to size40 so the decimal padding is consistent.
- ChordPicker's open-effect bound a window.setTimeout id to a local `t`, shadowing the i18n `t` from useTranslation. Renamed to timeoutId.
* perf(settings): persist generation sliders on release, not per pointer-move
Both sliders on the generation settings page were calling update() —
which is a React Query mutation that PATCHes /settings/generation —
inside onValueChange. Dragging the chunk-limit slider from 800 to 3000
fired a request per pointer-move pixel, and a mid-drag failure plus
optimistic rollback would leave persisted state visibly out of sync
with the thumb position.
Local state now mirrors each slider during a drag and the persist
happens once on Radix's onValueCommit (pointer-up / keyboard-release).
useEffects keep the local state in sync if the persisted value changes
out-of-band — another window editing the same setting still updates the
slider position cleanly.
* chore(backend): Ruff lint pass — deprecated APIs, exception leaks, dead patterns
Mechanical sweep of items called out in the PR review:
- qwen_llm_backend: AutoModelForCausalLM.from_pretrained(torch_dtype=…) is deprecated in transformers ≥4.41 in favor of dtype=. Renamed.
- routes/llm: try/except around backend.generate() raised HTTPException(500, detail=str(e)) which leaks stack traces / paths to clients and trips Ruff B904. Now logs the original exception server-side and hands the client a generic message; chained via `from e` to preserve traceback context.
- mcp_bindings + mcp_server/context: datetime.utcnow() is deprecated since 3.12. Switched the two assignment sites to datetime.now(timezone.utc). The schema-level `default=datetime.utcnow` defaults in database/models.py are left for a later schema-aware pass.
- routes/generations: `logger = …` sat between two import blocks (Ruff E402). Moved below imports.
- mcp_server/server + tests/test_refinement_samples: typing.Callable / typing.Iterable have been preferred-via collections.abc since 3.9 (Ruff UP035).
- routes/events: `except asyncio.TimeoutError` aliases plain `TimeoutError` since 3.11 (UP041).
- services/captures: hoisted WHISPER_NATIVE_FORMATS to module scope (was a function-local UPPER_SNAKE that tripped N806) and replaced the raw_path.unlink try/except OSError-pass with contextlib.suppress (SIM105). Semantic equivalence preserved — written_files.remove(raw_path) still only runs when unlink succeeds because it sits inside the suppressed block after the unlink call.
- database/migrations: hoisted the duplicate `import sqlite3` from inside two helper bodies to a single module-level import.
* feat(stories): regenerate action on clips and the chat list dropdown
The track editor's clip toolbar now has a regenerate icon next to Delete; clicking it kicks a fresh take of the selected clip's underlying generation through the same /generate/{id}/regenerate path the History table uses, and pushes the id into the global pending set so the SSE watcher picks it up. The chat list's per-item dropdown gets the same action between Play-from-here and Remove. Translation keys added under storyContent.itemActions / storyContent.toast across all four locales.
* feat(stories): import external audio into the timeline (drag-drop + picker)
You can now drop a music file onto the story content area or pick one through the new "Import audio" button in the add-clip popover. Both call POST /generate/import which writes the file to data/generations/<id>.<ext>, probes duration via librosa, and inserts a Generation row pointing at a singleton "Imported Audio" profile (created lazily on first import). The existing addStoryItem flow takes over from there — the timeline doesn't care that the row didn't come out of TTS.
Engine field on the row is "import"; it's surfaced on StoryItemDetail so the chat list shows a music icon instead of the (missing) profile avatar and both the dropdown and the track-editor toolbar hide the Regenerate action — there's nothing to regenerate. Accepted formats: wav/mp3/flac/ogg/m4a/aac/webm, capped at 200 MB. Translation keys added across en/ja/zh-CN/zh-TW.
* fix(audio): serve real Content-Type so imports decode in WaveSurfer
/audio/{id} and /audio/version/{id} hardcoded media_type="audio/wav" on
the FileResponse. That was a no-op when every generation came out of
TTS (everything on disk was a .wav anyway), but imported audio keeps
its source format — .mp3 / .m4a / .ogg — and the WaveSurfer MediaElement
backend uses an <audio> tag that checks Content-Type before letting the
clip play, so an MP3 announced as audio/wav silently failed to load.
Both endpoints now derive the type via mimetypes.guess_type and fall
back to audio/wav for unknown suffixes. Download filenames also keep
the real extension instead of always saying ".wav".
* feat(stories): zoom bar bounds tracked to project length, default 60s scope
The track editor's zoom was clamped to a hardcoded [10, 200] pixels-per-second range, which had no relationship to the project — on a 4-minute story a "max zoom out" of 200 px/s still required scrolling, and on a 5-second story you could zoom all the way in to where every clip was a tiny sliver. Reframed the bounds in the unit the user actually thinks in: how many seconds of timeline are visible at once. Min scope is 10 s (most zoomed in), max scope is the entire project, and the default lands on a 60 s scope (or the full project, whichever is shorter) once the editor measures its visible track width on first mount.
The pixels-per-second value still lives in component state (because every downstream calculation already uses it) but minPps/maxPps are computed from `containerWidth − LABEL_COL_WIDTH` and the project's effective duration, so the +/- buttons and the edge-drag handles on the scrollbar all clamp to bounds that move with the project. Re-clamping fires whenever those bounds shift — adding a long clip or resizing the window pulls the current zoom inside the new range instead of leaving the user parked outside it.
* fix(stories): show the source filename on imported clips
Imports were rendering as "Imported Audio" everywhere because every
import points at the singleton voice profile. The filename was already
being stored on the generation row (in the `text` field), so the chat
item title and the timeline clip label now read from `text` when
`engine === 'import'` and fall back to the profile name otherwise. The
chat item also drops the language pill (always "en" on imports — not
informative) and skips the transcript textarea since imports have no
spoken text to show.
* fix(stories): round split_time_ms before posting
handleSplit was sending currentTimeMs - item.start_time_ms straight to the backend, which rejects it because StoryItemSplit.split_time_ms is typed as int and the playhead's currentTimeMs is a float (it's driven from HTMLAudioElement.currentTime, which carries sub-millisecond precision). Pydantic surfaced the mismatch as "Input should be a valid integer, got a number with a fractional part" and the toast read "Failed to split clip". Math.round at the call site, matching what the trim and move handlers already do.
* feat(stories): per-clip volume control on the timeline
Each story item now carries a volume column (linear gain, default 1.0,
clamped 0.0–2.0 server-side). New PUT /stories/{}/items/{}/volume route
+ useUpdateStoryItemVolume hook + a Volume2 icon in the clip-edit
toolbar that opens a popover with a 0–200% slider. Local slider state
drives the visual during a drag; the persist fires once on
onValueCommit, mirroring the generation-page slider pattern.
Web Audio playback inserts a per-clip GainNode between source and
master so volume changes apply live without re-decoding the buffer
(source -> clipGain -> masterGain -> destination). Server-side
mixdown in export multiplies the trimmed clip by its volume before
summing into the timeline. Split + duplicate carry the volume forward
to the new clips so trimming a faded section keeps the level you set.
Migration adds the volume column with default 1.0 so existing rows
read as full volume.
* fix(stories): mute the clip waveform's media element so it can't bleed audio
The clip waveforms drawn inside each timeline track use WaveSurfer with the default MediaElement backend, which creates an internal <audio> element to drive playback timing. Web Audio in useStoryPlayback is what actually produces sound, but WaveSurfer's element was happily preloading and — after the first user gesture unlocked browser autoplay — playing the source URL through the page output too.
For TTS clips it was masked: they're short, both sources start at the same time, and stopping the BufferSourceNode at pause coincides with the natural end of the audio element. For long imports (a four-minute MP3) the BufferSourceNode stops on pause but WaveSurfer's element keeps going on its own track — which is exactly the "music keeps playing when I pause" symptom.
Hand WaveSurfer a muted <audio> element via the `media` option so the visual still loads peaks but the element itself can never produce sound. preload="metadata" keeps the load lightweight.
* fix(stories): hard-cut the audio graph on stop so long imports actually halt
source.stop() was the only thing happening when a clip was halted, and on long imported buffers (multi-minute MP3s scheduled via source.start with a duration argument) it was silently failing to halt the buffer in some browsers — pause left the music playing and seek stacked another source on top of the original. The mute-the-WaveSurfer-element fix was a different bug along the same path; this is the one that actually addresses the duplicated audio.
ActiveSource now carries the per-clip GainNode alongside the source, and stopSource detaches the onended handler before calling stop() (so the natural-end callback can't race with explicit teardown and re-delete a freshly rescheduled entry at the same id), then disconnects both nodes inside their own try/catch blocks. Even when stop() doesn't actually halt the buffer the graph is severed — no path from source to destination, no audio.
* feat(stories): add empty tracks above/below the timeline
Tiny + strips sit at the top of the topmost label cell and the bottom of the bottommost one, sticky-positioned in the label column so they follow horizontal scroll. Clicking either extends the visible track stack in that direction by one — above adds max(existing)+1, below adds min(existing)-1. Both compute against the full set (defaults + item-derived + previously-added) so successive clicks keep extending instead of fighting over the same number.
Empty extras live in component state because a track only earns its keep once a clip lands on it. Once one does, item.track carries the number forward and the row keeps deriving from items naturally; if nothing lands there before reload, the empty row simply isn't there next time, which matches what the user expects of an unused affordance.
* fix(mcp): bundle stdio shim sidecar
* fix(captures): allow dictation without paste permission
* fix(mcp): preserve speak engine defaults
* fix(captures): use platform hotkey defaults
* fix(mcp): preload speak pill window
* fix(captures): hide unwired storage settings
* feat(sponsors): add /sponsors page, homepage promo, and in-app strip
* style(landing): drop pill chrome from /download maintainer kicker
* changelog
* better naming for sponsors
* windows keybind note
---------
Co-authored-by: Claude Opus 4.7 (1M context) <[email protected]>
fugashi (pulled in by misaki[ja]) needs a MeCab dictionary at runtime.
The `unidic` package that ships today contains no data — it relies on
`python -m unidic download` (~526MB), which isn't run by `just setup`
and won't survive PyInstaller freezing.
Switch to `unidic-lite`, which bundles a MeCab-compatible dict inside
the wheel (~50MB). Collect its data files in build_binary.py so frozen
builds also pick up the dicdir. Same failure mode and same fix shape as
the existing en_core_web_sm pre-install.
* fix(build): bundle kokoro source files for transformers runtime introspection
transformers opens .py source files at runtime to check attention/MoE
implementation via regex (e.g. _can_set_attn_implementation). PyInstaller's
--hidden-import only bundles .pyc bytecode, so kokoro/modules.py was missing
from the bundle causing a FileNotFoundError on Kokoro model load.
Switch from individual --hidden-import entries to --collect-all kokoro in both
build_binary.py and voicebox-server.spec. The kokoro package is 172K so no
meaningful bundle size impact.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
* fix(build): use SPECPATH for runtime hook instead of hardcoded absolute path
The linter expanded runtime_hooks=[] to an absolute /Users/... path which
would break CI and other dev machines. Use os.path.join(SPECPATH, ...) to
mirror the relative approach in build_binary.py.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
* fix(build): runtime hook to work around PyInstaller + Python 3.12 import breakages
Four distinct bundling-specific crashes blocked Kokoro and Qwen CustomVoice
from loading in the frozen binary:
1. torch._dynamo import triggered via class-body decorators
(@torch._dynamo.allow_in_graph on PreTrainedModel,
@torch.compiler.disable in flex_attention) pulls in torch._numpy._ufuncs
which crashes on module load with NameError: name 'name' is not defined.
2. AlbertModel (Kokoro) triggers @auto_docstring -> modeling_auto ->
GenerationMixin -> candidate_generator -> sklearn -> scipy, which hits
the same class of bug in scipy.stats._distn_infrastructure (NameError:
name 'obj' is not defined).
3. AutoModel (Qwen) pulls the same sklearn -> scipy chain directly.
4. librosa (required by most TTS engines) -> scipy.signal -> scipy.stats
hits the _distn_infrastructure crash regardless of the transformers
stubs above.
The root cause of (1) and (4) is that PyInstaller's frozen importer runs
module-level `for X in [<list-comp using dir()>]:` loops with an empty
iterable, leaving the loop variable unbound. Trailing `del obj` / unrelated
references then crash.
Fix: a single runtime hook (pyi_rth_torch_compiler_disable.py) installs:
- sys.modules stubs for torch._dynamo and torch._dynamo.config, plus a
meta-path finder for torch._dynamo.* submodules — voicebox never uses
torch.compile/dynamo for inference, so a permissive no-op stub (callable
as decorator, falsey as predicate, context-manager-safe for
TransformGetItemToIndex) is drop-in safe.
- meta-path finder stubs for transformers.utils.auto_docstring and
transformers.generation.candidate_generator — both import-chain
short-circuits; docstrings and speculative decoding aren't used for TTS.
- meta-path finder for scipy.stats._distn_infrastructure that reads the
real .py source via the wrapped loader's get_source(), replaces the
bundling-broken `del obj` with `globals().pop('obj', None)`, and
compile+exec's the patched source. This keeps the real scipy module
intact so librosa and everything downstream works normally.
Supporting changes:
- backend/pyi_hooks/hook-scipy.stats._distn_infrastructure.py sets
module_collection_mode = "pyz+py" so the .py source is actually in the
bundle for the runtime patcher to read.
- build_binary.py and voicebox-server.spec register the runtime hook and
the new hooks dir.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
* fix(build): force transformers torch<2.6 mask path and bundle spacy_pkuseg
- patch transformers.masking_utils to set _is_torch_greater_or_equal_than_2_6
= False, forcing sdpa_mask_older_torch and avoiding the vmap .item() crash
that breaks Qwen CustomVoice generation (our torch._dynamo stub can't
reproduce TransformGetItemToIndex's graph transform).
- add PyInstaller hook to bundle transformers.masking_utils .py source so the
runtime finder can source-patch it.
- --collect-all spacy_pkuseg so Chatterbox Multilingual can load its Chinese
segmenter (dicts/default.pkl + native .so extensions).
- add per-finder install diagnostics + _HOOK_VERSION marker to make future
bundle-only regressions easier to triage.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(build): pass PyInstaller hook paths relative so .spec is portable
Absolute paths ended up in the auto-regenerated voicebox-server.spec
because build_binary.py prefixed every --runtime-hook and
--additional-hooks-dir with str(backend_dir / ...). That broke builds
on any machine whose checkout wasn't at /Users/jamie/... and anyone
invoking pyinstaller voicebox-server.spec directly.
os.chdir(backend_dir) already runs before PyInstaller (same reason
server.py works as a bare filename), so the backend_dir prefix is
unnecessary. Drop it so the generated spec references pyi_hooks/,
pyi_rth_numpy_compat.py, pyi_rth_torch_compiler_disable.py as repo-
relative paths.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>
torch is compiled against numpy 1.x. numpy 2.x changed the ABI version
returned by PyArray_GetNDArrayCVersion() (0x01000009 → 0x02000000), so
torch's is_numpy_available() always returns False and torch.from_numpy()
raises RuntimeError. This causes TTS generation to fail with:
ValueError: Unable to create tensor, you should probably activate
padding with 'padding=True'
Two fixes:
1. Pin numpy<2.0 in requirements.txt so new builds bundle a compatible
numpy version. (The existing comment already flagged this intention
but the upper bound was never added.)
2. Add a PyInstaller runtime hook (pyi_rth_numpy_compat.py) that installs
a ctypes memmove fallback for torch.from_numpy() at startup. Runtime
hooks run after FrozenImporter is registered so frozen torch is
importable. The fallback catches RuntimeError from the C-level ABI
check and copies the numpy array into a new tensor via raw memory copy,
bypassing the check entirely. This is a belt-and-suspenders fix that
works regardless of the bundled numpy version.
Co-authored-by: aimaaaimaa <[email protected]>
Co-authored-by: Claude Sonnet 4.6 <[email protected]>
Add Kokoro-82M as a new TTS engine — 82M params, CPU realtime, 8 languages,
Apache 2.0. Unlike cloning engines, Kokoro uses pre-built voice styles, which
required a new profile type system to support non-cloning engines cleanly.
Kokoro engine:
- New kokoro_backend.py implementing TTSBackend protocol
- 50 built-in voices across en/es/fr/hi/it/pt/ja/zh
- KPipeline API with language-aware G2P routing via misaki
- PyInstaller bundling for misaki, language_tags, espeakng_loader, en_core_web_sm
Voice profile type system:
- New voice_type column: 'cloned' | 'preset' | 'designed' (future)
- Preset profiles store engine + voice ID instead of audio samples
- default_engine field on profiles — auto-selects engine on profile pick
- Create Voice dialog: toggle between 'Clone from audio' and 'Built-in voice'
- Edit dialog shows preset voice info instead of sample list for preset profiles
- Engine selector locks to preset engine when preset profile is selected
- Profile grid filters by engine — shows Kokoro voices when Kokoro selected
- Custom empty state when no preset profiles exist for selected engine
Bug fixes:
- Fix relative audio paths in DB causing 404s in production builds
- config.set_data_dir() now resolves to absolute paths
- Startup migration converts existing relative paths to absolute
Also updates PROJECT_STATUS.md and tts-engines.mdx developer guide.
Upgrade CUDA toolkit from 12.6 (cu126) to 12.8 (cu128) for proper
RTX 50-series (Blackwell) GPU support. Users with RTX 5070/5080/5090
were reporting CUDA detection failures with cu126.
Also fix the GPU Acceleration settings panel where the 'Switch to CPU
Backend' button was unreachable — it was inside a conditional block
that required !isCurrentlyCuda, making it impossible to switch back
to CPU once running on CUDA.
Closes#315
Replace --collect-submodules + --collect-data with --collect-all for
qwen_tts. The qwen_tts runtime expects physical .py source files
(e.g. modeling_qwen3_tts.py) under _MEIPASS, which only --collect-all
provides. This is the same pattern used for inflect/typeguard.
Fixes#212
Switch CUDA builds from PyInstaller --onefile to --onedir and split the
output into two separately versioned archives:
1. Server core (~200-400MB) — versioned with the app, redownloaded on
every app update
2. CUDA libs (~2GB) — versioned independently (cu126-v1), only
redownloaded when the CUDA toolkit or torch version changes
This eliminates the ~2.4GB full redownload on every version bump.
After initial setup, most app updates only need ~200-400MB.
Closes#297
The real descript-audio-codec package pulls in descript-audiotools,
which transitively requires onnx, tensorboard, protobuf, matplotlib,
pystoi, and other heavy dependencies. onnx fails to build from source
on macOS due to CMake version incompatibility.
TADA only uses Snake1d (a 7-line PyTorch module) from DAC. This commit
adds a shim in backend/utils/dac_shim.py that registers fake dac.*
modules in sys.modules with just the Snake1d class, completely
eliminating the DAC/audiotools dependency chain.
Integrates HumeAI's TADA (Text-Acoustic Dual Alignment) speech-language
model as a new TTS engine. TADA uses a novel 1:1 token-audio alignment
that produces coherent speech over long sequences (700s+).
Two model variants:
- tada-1b: English-only, ~4GB, built on Llama 3.2 1B
- tada-3b-ml: 10 languages, ~8GB, built on Llama 3.2 3B
Backend uses the Encoder for voice prompt encoding with caching, and
TadaForCausalLM with flow-matching diffusion for generation. Supports
bf16 inference on CUDA, forces CPU on macOS (MPS compatibility).
Installed with --no-deps due to torch>=2.7 pin conflict; descript-audio-codec
and torchaudio added as explicit sub-dependencies.
- Rename Server tab to Settings with horizontal sub-tab navigation (General, Generation, GPU, Logs, Changelog)
- All sub-tabs are proper routes under /settings/* with /server redirect for backwards compat
- General: connection settings, link cards (docs + discord), API reference card, app updates
- Generation: auto-chunking, crossfade, normalize, autoplay as SettingRow components
- GPU: info card with platform-aware icons (Apple logo for MPS), CUDA management, explainer text
- Logs: real-time server log viewer piped from Tauri sidecar via event system (Tauri-only)
- Changelog: parsed from CHANGELOG.md at build time via Vite virtual module plugin
- New reusable SettingRow/SettingSection components for consistent settings layout
- New Toggle (switch) UI component replacing checkboxes in settings
- Toast viewport now offsets when audio player is open
- Sidebar stays active on settings sub-routes (fuzzy matching)
Replace verbose startup messages with a clean summary:
- App version, Python version, OS/arch
- Database path (fix None display), data directory
- Profile and generation counts
- Backend, GPU, model cache path
- Clean up stale loading_model status on startup
- Remove noisy progress manager log line
- Use DB COUNT query instead of list length for take-N label to avoid
TOCTOU race between list_versions and create_version
- Add focus:bg-muted to SelectTrigger for keyboard focus visibility
Critical:
- Remove dead backend.utils.validation PyInstaller hidden import
- Fix story_items table rebuild to preserve track/trim/version columns
- Guard cache migration against same source/destination path
- Fix regeneration audio overwrite (use random uuid suffix per take)
Major:
- Engine selector: validate language on Qwen switch, clear stale modelSize
- Sync language validation regex between profile create and generate (22 langs)
- Guard CUDA download against duplicate concurrent requests
- Only set model_size for engines that support multiple sizes
- Fix 404 swallowed by generic except in history export
- Validate audio_path before FileResponse in export-audio
- Transcription: stream uploads in 1MB chunks, use robust cache check,
call complete_download() on Whisper download success
- Set clean version as default when effects chain validation fails
- Return explicit error when Windows port occupied by non-voicebox process
Split the 2,578-line main.py (90 routes) into 12 domain-specific router
modules under routes/. main.py is now a 45-line entry point.
New structure:
- app.py: FastAPI instance, CORS, startup/shutdown, safe_content_disposition
- routes/: health, profiles, channels, generations, history, transcription,
stories, effects, audio, models, tasks, cuda
- services/cuda.py: moved from cuda_download.py
Also includes Phase 5 database/ package (from parallel agent):
- database/__init__.py re-exports all symbols for backward compat
- database/models.py, session.py, migrations.py, seed.py
All 90 routes verified registered and app imports cleanly.
- collect-all piper_phonemize to bundle espeak-ng-data for LuxTTS phonemization
- Set ESPEAK_DATA_PATH in frozen builds so the C library finds bundled data
- collect-all perth to bundle pretrained watermark model for Chatterbox
- Add multiprocessing.freeze_support() to fix resource_tracker subprocess crash
Add the ability to download a CUDA-enabled backend binary (~2.4 GB) and
swap it in via a backend-only restart, solving the #1 user pain point
(19 open 'GPU not detected' issues caused by GitHub's 2 GB asset limit).
Backend:
- cuda_download.py: download from R2 (primary) or GitHub split-parts
(fallback), SHA-256 verification, atomic writes, progress via SSE
- 4 new endpoints: GET/POST/DELETE /backend/cuda-*, GET cuda-progress
- server.py: --version flag, auto-detect variant from binary name
- build_binary.py: --cuda flag for CUDA PyInstaller builds
- split_binary.py: split large binaries into <2GB GitHub Release assets
- CI workflow for building CUDA binary
Tauri:
- restart_server command (stop -> wait -> start)
- start_server prefers CUDA binary from {data_dir}/backends/ if present
- Version mismatch check: runs --version before launching CUDA binary
Frontend:
- GpuAcceleration component: download, progress, restart, switch, delete
- API client + types for CUDA status and management
- Platform lifecycle: restartServer() on Tauri/Web
- Aggressive 1s health polling during restart for fast reconnection
The distributed macOS aarch64 binary shipped without MLX acceleration despite
the model and backend code supporting it. Two root causes:
1. **OSError not caught in platform_detect.py**
PyInstaller bundles isolate the filesystem, so when MLX tries to load its
Metal shader libraries (.metallib) it raises OSError, not ImportError.
platform_detect.get_backend_type() only caught ImportError, causing a
silent fallback to PyTorch even on Apple Silicon hardware.
Fix: broaden the except clause to (ImportError, OSError, RuntimeError)
and import mlx.core instead of mlx (forces native lib loading eagerly).
2. **collect_data_files used instead of collect_all for MLX**
build_binary.py and voicebox-server.spec used --collect-data /
collect_data_files for mlx and mlx_audio. This copies Python source and
pure-Python data, but NOT native shared libraries (.dylib, .metallib).
Fix: switch to --collect-all / collect_all which captures binaries too,
then pass them to Analysis(binaries=...) in the spec.
Result: macOS Apple Silicon users now get MLX inference (~4-5x faster than
PyTorch CPU), matching the performance documented in the README.
- Updated hidden imports in build_binary.py to replace 'mlx_audio.asr' with 'mlx_audio.stt'.
- Enhanced model loading logic in MLX and PyTorch backends to ensure proper progress tracking during model downloads.
- Improved error handling and context management for progress tracking in both backends.
- Bumped version to 0.1.10 in Cargo.lock to reflect recent changes.
- Added support for MLX backend on Apple Silicon, enabling optimized performance for TTS and STT tasks.
- Implemented platform detection to dynamically select between MLX and PyTorch based on the runtime environment.
- Updated build process to include MLX-specific dependencies and configurations for macOS.
- Refactored backend code to improve model loading and inference logic, accommodating backend-specific requirements.
- Enhanced documentation to clarify backend selection and performance benefits for different platforms.
- Streamlined installation instructions and troubleshooting guidance for MLX-related issues.
- Added support for MLX backend on Apple Silicon, enabling optimized performance for TTS and STT tasks.
- Updated release workflow to include MLX-specific dependencies and configurations for macOS platforms.
- Refactored backend code to dynamically select between MLX and PyTorch based on the runtime environment.
- Enhanced model loading and inference logic to accommodate backend-specific requirements, including updated model IDs and hidden imports.
- Improved health check and model status reporting to reflect the active backend type.
- Streamlined caching mechanisms to support both backend types, ensuring compatibility and performance.
- Updated CONTRIBUTING.md to include instructions for building with a local Qwen3-TTS development version, facilitating easier testing and development.
- Refactored FloatingGenerateBox component to streamline the rendering of text and instruct fields, improving code readability and maintainability.
- Added functionality to handle auto-resizing of text areas based on content changes, enhancing user experience.
- Improved event handling for keyboard interactions in StoryTrackEditor, allowing for play/pause functionality with the spacebar.
- Introduced a MiniSamplePlayer component in SampleList for better audio playback control, including play, pause, and seek features.
- Implemented sample update functionality in the backend, allowing users to edit reference text for audio samples, with appropriate error handling and user feedback.
- Added exclusions for 'torch.utils.tensorboard', 'scipy', 'PIL', 'tkinter', 'unittest', and 'test' to reduce bundle size and improve build efficiency.
- Install llvm-dev on Ubuntu for llvmlite compilation
- Install LLVM via Homebrew on macOS and configure PATH
- Exclude matplotlib, IPython, notebook, pytest, tensorboard from bundle to reduce size
- Keep --onefile mode with module exclusions to avoid 4GB limit
