[joshfrenette22-stack]

What

A searchable voice picker in the generate box, with a ▶ button on every voice that plays a short sample phrase before you generate. Built-in Kokoro and Qwen CustomVoice voices are listed alongside your own voices, so you can audition them without leaving the generate box.

Also fixes a Kokoro bug that made every voice you hadn't used before fail on generation (second commit below).

Why

Right now the generate box has no way to choose a voice (only the Stories tab had a plain dropdown), and the Create Voice dialog's built-in voice grid can't be auditioned. You pick a name from a list of 54 Kokoro voices blind, create a profile, and only then hear it.

How it works

Picker (VoicePicker.tsx)

• A pill in the generate box's bottom row opens a popover with a search field, a "My voices" section, and a section per built-in engine (Kokoro, Qwen CustomVoice).
• Choosing one of your voices selects it. The existing profile-sync effect then switches engine and language.
• Choosing a built-in voice reuses the matching preset profile if there is one, otherwise creates it (name de-duplicated, since profile names are unique) and selects it. Generation is always tied to a profile, so this avoids a separate "add voice" step.
• Replaces the plain voice Select that only the Stories tab had, so the showVoiceSelector prop is gone.

Preview playback (useVoicePreview, VoicePreviewButton)

• One clip plays at a time. Clips are cached as object URLs for the life of the hook, so replays are instant, and a slow response for an earlier click won't talk over a newer one.
• The Create Voice dialog's built-in grid gets the same buttons.

Backend

• GET /profiles/presets/{engine}/{voice_id}/preview: renders a short phrase in the voice's own language (11 phrases; unknown languages fall back to English) with a fixed seed and caches the WAV under the cache dir. Cache file names are digests, so a voice id never shapes a path.
• GET /profiles/{profile_id}/preview: same for any profile. Built-in profiles share the preset cache; cloned and designed voices render on demand (not disk-cached, since editing the profile changes the sound).
• Previews run one at a time behind a lock, since they share the model with real generations.
• A model that isn't downloaded returns a 409 pointing at the Models tab, instead of the stock message that refers to /generate.

Kokoro fix (separate commit)

fix(kokoro): the model download only fetches the weights; each voice is a separate file that KPipeline pulls from the Hub on first use. huggingface_hub keeps one HTTP session per thread, and a session built during an offline window (opened while loading a cached model) keeps its offline adapter. A thread reused for generation then fails with OfflineModeIsEnabled for every voice that hasn't been used before. This reproduced on a plain generation, not just previews. The fix checks the voice file is cached and, if not, fetches it inside a network-allowed window with fresh sessions.

Testing

• New backend tests (13): phrase coverage for every built-in voice language, cache path stability, render-once-then-cache, 404s, the missing-model message, and the Kokoro voice fetch. The renderer and Hub are stubbed, so no model is loaded.
• tsc passes for app and web. Biome reports nothing in the new files (remaining findings are in lines I didn't touch).
• Run against a real backend:
  • Kokoro previews for voices that weren't cached (af_bella, bm_george) rendered and downloaded the voice on demand. Transcribing the clips with the app's Whisper returned the exact phrase.
  • A cloned profile (Qwen base) previewed end to end and transcribed back to the phrase; the throwaway profile was deleted afterwards.
  • Opened the picker in the app and checked preview, selecting a built-in voice (profile created, engine and language switched), and search.
  • In the Create Voice dialog's built-in grid, ▶ plays a voice (including one whose file wasn't downloaded yet) without selecting its card.

Not covered / notes

• UI checks were done in a browser against the Vite dev server, not the packaged Tauri build, and there are no screenshots attached.
• Qwen CustomVoice previews only had the "model not downloaded" (409) path exercised, and designed-voice previews weren't run (no VoiceDesign model on hand).
• Only English strings were added; other locales fall back to English.
• backend/tests/test_offline_guard.py::test_nesting_opposite_mode_on_same_thread_raises_instead_of_deadlocking fails on current main too, independent of this change.

🤖 Generated with Claude Code