New Downloadable Voices: Using MOSS-TTS-Nano on the Web

The web app now offers MOSS-TTS-Nano, a downloadable speech engine for desktop browsers. It replaces the previous English-only SpeechT5 HD option with voices for 19 languages. Once installed, it generates speech on your computer rather than sending your card text to a speech server.

MOSS is an open-source model distributed under the Apache-2.0 license. The download is optional and about 733 MB, including the runtime and license notices. Opening a study screen does not start a model download.

Download and choose your speech engine

1. Open Settings → Text-to-Speech in the web app on a supported desktop browser.
2. Find Natural voices, check the displayed download size, and select Download.
3. Wait for Downloaded to appear. You can use Cancel while the download is in progress.
4. Under Speech engine, choose MOSS-TTS-Nano.
5. Choose Female voice or Male voice, then use Listen to sample to check playback.
6. Turn on Enable Text-to-Speech for spoken cards. Enable Auto-play Audio if you want audio to start when a card appears in card review. Audio Review has its own explicit Play/Pause controls; see the player walkthrough.

Check your set's Front Language and Back Language in its editor. These tell the speech system which language each side uses; changing the interface language is a separate setting.

Which languages use the downloaded voices?

MOSS supports Arabic, Chinese, Czech, Danish, English, French, German, Greek, Hungarian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Russian, Spanish, Swedish, and Turkish.

Other languages use an installed device voice. Downloading MOSS does not add a voice for every language in the app. For a language outside this list, check which voices your browser and operating system provide.

You can choose Device voices (browser / operating system) under Speech engine at any time. That keeps the downloaded MOSS pack, so you can switch back without downloading it again. Remove download deletes the pack if you want to reclaim its storage.

Allow time for preparation

New text and the first use of a model can take time to prepare. MOSS reuses its loaded voice setup and saves generated audio locally so that repeat playback, including after a reload, can reuse it. In Audio Review, it also prepares one upcoming utterance. This reduces repeated work; it does not make every new card instant or guarantee uninterrupted playback.

The saved audio cache is bounded to 64 MB, so older audio may need to be generated again. Remove download clears the model pack and its saved audio. The saved audio remains on this computer in the browser profile where it was generated.

Speed depends on your computer, browser, and text length. Firefox was substantially slower than Chromium in our Linux checks. Start with a short sample; if preparation feels slow, try a Chromium-based desktop browser and compare it there.

If playback reports a problem, allow site audio and check the computer's sound output. Switching to device voices is another option, provided a suitable device voice is installed.

Local processing and cached downloads

An installed engine can restart offline from its browser cache. After an app runtime update, its first use may need an internet connection to cache the updated runtime; installed model weights are retained.

This describes the speech engine, not a promise that the website and every deck are available offline. The download belongs to the browser profile where you installed it. Another browser or profile needs its own copy, and clearing site data can remove both local app data and the model cache.

For a complete language-card workflow, see translation and speech together. To generate draft answers in the editor, see local M2M100 translation.