Inside the project
Prepared with the project’s development agent from source and retained evidence. This is a project-owner assessment, not an independent customer review. Dated checks were not rerun for this website.
A family of on-device voice tools inside AssistantApp for macOS. You make a personal neural voice with explicit consent, either by recording guided prompts or by importing a performance you have the rights to. You can then derive variations of it (deeper, brighter, breathier, other languages) and pick any of these voices in the app's narration modules: Audiobook Studio, Actor/Script Studio, Conference Theatre, NeuroAtlas narration and spoken chat replies. Rendering runs locally on Apple silicon through an open-source (MIT) Chatterbox model under MLX, with a graceful fallback to Apple system voices. It is a working research prototype. Expression follows measured delivery (energy, pace, stability), but naturalness depends on the reference recording. Rendering is not real-time: a warm render takes a few seconds per sentence.
What stands out
- Consent and rights confirmation gate every voice creation and every training run
- Import any audio format: the performance is transcribed, split into phrases at the speaker's pauses, and prosody markers (emphasis, breaks, pace, arousal) are measured into an expression timeline
- One shared voice catalogue, so neural, designed and Voice Studio voices appear in every narration module's picker
- Voice Designer: pitch-preserving formant (throat) warp, pitch shift, breath, brightness and tension, with presets; validated numerically to track target formant scaling
- Local hi-fi cloning (Chatterbox multilingual via MLX), offline at render time; a persistent helper cuts a warm render from about 107 s to about 7 s
- Delivery controls change the result: the same text measured 2.92 s calm vs 2.60 s excited; French and German verified in a cloned voice
- Honest labelling: the voiceprint replay is labelled as a replay of the captured clip, not new speech
Current limits
- Not real-time; the first (cold) render is slow, and the earlier toy diffusion path is very slow and prototype-grade
- Naturalness and timbre match depend on a clean reference of about 10–12 s; an ear-based quality judgement is the owner's alone, with no listening panel
- Timbre processing cannot create a genuine accent; accents rely on multilingual synthesis or accent-native Apple voices
- The model files are large (2.5 GB multilingual, kept on an external drive); with the drive unplugged it falls back to a smaller local model
- Correctness was checked by machine transcription (Whisper-tiny), not by listener ratings
- macOS only; the cross-platform build is not verified
- No released public build; this is a local development app
Features & evidence
Expressive Neural Voice panel
installed-testedSix steps: consent, add or import audio, calibrate, build, style and preview, publish as the app-wide voice
Recorded evidence
Full Xcode clean build succeeded 2026-08-31 with 0 errors; Apple-standard form redesign and import bar done August 2026
Voice Studio import digest
component-testedImport a performance; it is transcribed and phrased by pauses into a prosody-marked transcript, then becomes a reusable voice
Recorded evidence
Swift typecheck against the compiled kit passed, plus a logic check of the phrase grouper (Aug 2026); included in the 2026-08-31 clean build
Hi-fi clone renderer
component-testedSays any scripted text in the enrolled voice, with delivery and language controls
Recorded evidence
Regression suite passed (2026-08-31): scripted text renders through the clone, an exact training passage replays the real take, a broken toy model still renders hi-fi, and forced diffusion and forced source-backed modes work; Whisper transcript ≈ target text; FR/DE verified
'Plays a recording' fix
component-testedScripted text is now synthesised instead of silently replaying the closest training take
Recorded evidence
Verified 2026-08 on the owner's own 9-record dataset: an unrelated script scored 0.20 against the 0.80 threshold and was synthesised
Voice Designer
component-testedDerive named variations of a voice with sliders, presets and an accent voice, save them, and preview
Recorded evidence
Formant scale accuracy measured (0.75→0.74, 1.30→1.30), pitch shift duration-preserved; included in the 2026-08-31 build
Voice Setup guided flow
installed-testedPersonal-Voice-style flow: consent, a room check, read prompts with take verification, pitch sweeps, build; it no longer reports success falsely
Recorded evidence
Audit-sweep fixes (10 verified) shipped in the 2026-08-31 build
App voice catalogue
installed-testedNeural, designed, Voice Studio and Apple voices selectable everywhere; optional spoken chat replies
Recorded evidence
2026-08-31 build; adopted in Audiobook, Actor, NeuroAtlas and Conference
Per-singer neural re-singing (AA Music, related)
component-testedRe-sing one separated vocal part in another voice, placed back where that singer stood
Recorded evidence
Seed-VC end-to-end test passed 2026-10-16 on synthetic voices only; the UI is not built
Review updates & corrections
- No material change to these speech modules since 17 Oct
- Related: AA Music gained a source-tested per-part re-singing backend on 16 Oct (synthetic test only; no UI)
- Consent was confirmed by the owner on 2026-10-13 for the dataset records; one imported record is another speaker and must never be used in demos
Workflow walkthrough
A guide to the native workflow. These steps describe the app; only the recordings and images embedded on this page are available as website demonstrations.
- Open Creative → Expressive Neural Voice, then confirm consent and rights
- Record the guided prompts, or import a clean performance you have rights to; check the measured expression timeline
- Build the voice and publish it as the app-wide voice
- Type a new script in Style & preview and render it calm, then excited
- Open Voice Designer, derive a 'Deep' variation, save it, then pick it in Audiobook Studio and render a paragraph
Planned captures · not yet available
- Proposed capture: consent-gated voice creation and a scripted render, using a consenting actor or the owner's own voice
- Proposed A/B: the same line rendered calm vs excited, and a Voice Designer variation
- Proposed screenshot: Voice Designer sliders and presets (no voice audio)