This is a separate Windows SAPI 5 integration project for Kyutai Pocket TTS. Version 0.1.1 is a published evaluation release using the official 6.9M-parameter ReFL cloning model (66.5 MB weights) selected by the written specification and local model benchmark. The installer supports an English default voice and optional language packs. It warns before stopping the Pocket TTS worker and closing apps Windows reports are using Pocket TTS files. Voice Manager registers imported voices machine-wide; voice-state files remain in the creating user's profile, so multi-user behavior needs a permissions design. Edge Read Aloud word highlighting and NVDA voice switching with NVDA's WASAPI option off were verified on the development machine. Follow-up validation items, including clean-profile setup, offline-network operation, and broader NVDA navigation, are documented in the compatibility reports. No model was modified.
The target is selectable, offline English voices in Windows applications, with Pocket TTS audio-prompt voice cloning preserved as a supported utility. Paradee-SAPI informs Windows integration and verification only. Its model, performance, and footprint are not Pocket TTS claims.
English is included in the base installer. The other languages are optional checkboxes; each selected pack downloads its language-specific model, tokenizer, and default demo voice during setup. Installed packs work offline.
| Language | Default voice | Install option | Approx. pack download |
|---|---|---|---|
| English | Alba | Included | Included in base installer |
| French | Estelle | Optional | 227 MB |
| German | Juergen | Optional | 225 MB |
| Italian | Giovanni | Optional | 223 MB |
| Spanish | Lola | Optional | 225 MB |
| Brazilian Portuguese | Rafael | Optional | 226 MB |
| Dutch | Daan | Optional | 223 MB |
This repository is a source and documentation backup. It does not include installer binaries, downloaded model weights, generated voice states, source prompt audio, raw benchmark recordings, or local caches. Download the published installer from the Pocket TTS for Windows 0.1.1 release; optional language packs download during setup. Review the license audit for attribution and model terms.
-
Download and run PocketTTS-SAPI-Setup.exe. Setup requires administrator approval.
-
Accept the license screen. English: Alba (6.9M) is included. Check any optional language packs you want; each selected pack downloads about 220–227 MB during setup and then works offline.
-
Setup creates both a Start menu shortcut (Pocket TTS → Voice Manager) and a Pocket TTS Voice Manager desktop shortcut. Open either shortcut to clone a voice from an audio prompt. Or open PowerShell and run:
& "$env:ProgramFiles\PocketTTS\tools\VoiceManager.ps1"
Enter a voice name, choose an audio prompt, confirm that you are authorized to use it, and select Create and register voice. The new voice is registered for Windows SAPI applications.
-
Fully exit and reopen Edge to refresh its Read Aloud voice list. In NVDA, choose Microsoft Speech API version 5 (64 bit) and leave Use modern audio output system (WASAPI) unchecked for Pocket voices.
The updated Voice Manager adds a Windows default voice dropdown. Select an installed voice and choose Preview choice to hear it before changing defaults. Set as default changes this user's OneCore default and, when the selected voice has a matching token, the classic SAPI default too. Restore previous puts both prior settings back. These are per-user settings and do not require administrator approval. The published 0.1.1 installer predates this control. If you already have Pocket TTS installed and a project checkout, run Run-PocketTTS-Voice-Manager.cmd from the checkout to open the updated manager against the existing install; this does not reinstall Pocket TTS.
To use Pocket TTS in Chrome Reading mode, set Alba (Pocket TTS) as the Windows default with Voice Manager, fully exit and reopen Chrome, then select System text-to-speech voice under Reading mode's Settings → Voice selection. Chrome does not list Alba by name there; the system option uses the Windows default. This flow was confirmed on the development machine. The Google US/UK Natural choices are Chrome-provided voices. Other Windows apps may use OneCore, classic SAPI, Narrator settings, or an app-specific voice choice, so restart them after changing a default.
The English model and installed voices run offline. The published Setup installer is unsigned; its SHA-256 is listed in the release notes.
Latest compatibility and latency evidence is recorded in Windows app compatibility, NVDA voice benchmark, SAPI latency benchmark, and SAPI rate benchmark. SAPI rate control selects slow, normal, and fast cached voice states; classic SAPI rate and XML tests pass. The latest local proof passed 20 rapid stop/restart trials per rate band. Edge Read Aloud highlighting passed. In NVDA 2026.3.1, Pocket voice selection fell back to Alba with "Use modern audio output system (WASAPI)" enabled; turning that NVDA option off allowed cloned and language tokens to reach the engine. The WASAPI-on path and NVDA navigation latency remain open.
Project handoff and conversation link: notes.md.
Pocket TTS is created by Kyutai. This is an independent Windows SAPI adapter and is not made or endorsed by Kyutai. The runtime is MIT licensed; model weights, tokenizers, and bundled demo voices carry their own CC BY 4.0 attribution and license terms. The included English Alba voice is attributed to Alba MacKenna. SoundTouch, by Olli Parviainen and contributors, is used under LGPL-2.1. Full license texts, asset attributions, source revisions, and change notes are included in C:\Program Files\PocketTTS\LICENSES after installation. Paradee-SAPI was consulted as a separate Windows integration reference; its code and model are not part of this project.
