SentenceSpeak for JLab decks: Japanese audio + Vietnamese translation

SentenceSpeak turns Japanese sentence cards—especially JLab-style cards—into guided listening lessons. Press one key while reviewing to hear the Japanese sentence clearly, followed by a natural Vietnamese translation and a short explanation of the speaker’s intent.

You can also select text anywhere in Anki and translate it instantly in a popup.

WHAT IT DOES

1. Read the current card aloud — press K while reviewing

   • Immediately stops the card’s original audio.
   • Plays a short opening chime.
   • Reads the Japanese sentence again using a clear Microsoft voice at normal speed.
   • Reads a natural Vietnamese translation (A), written in a fansub-like style with context-appropriate pronouns, implied subjects and sentence-ending nuances.
   • Reads one short explanation (B): what the speaker feels, wants or implies, based on the English meaning and scene notes stored in the card.
   • Plays a closing chime when the lesson finishes.
   • Press K again to stop and restart the current lesson.
   • Moving to another card immediately stops any remaining audio from the previous card.

Translation and speech generation start only after you press K. Simply viewing or skipping a card does not consume AI tokens.

2. Translate anything — Ctrl+Shift+T, macOS ⌘⇧T

   • Select text anywhere in Anki and press the shortcut, or right-click ▸ Translate with AI.
   • A popup next to the cursor translates the selected text while preserving line structure.
   • Click Explain for natural meaning, situational context, vocabulary with readings, grammar and nuance.
   • Click Copy to place the result on the clipboard.

SETUP

1. Create an account at https://console.river.ai and obtain an API key beginning with rv_.

2. In Anki, open:

   Tools ▸ SentenceSpeak ▸ Settings… ▸ AI

   Paste the key and click Test API.

3. Start reviewing and press K.

SentenceSpeak automatically installs a small Microsoft TTS client of approximately 5–10 MB. It does not download large voice models and does not require an Azure account or speech API key.

CARD FORMAT

By default, SentenceSpeak uses:

• Jlab-Kanji: Japanese sentence
• RemarksBack: English meaning and scene notes

For other note types, change the field names under Settings ▸ Read Aloud. Common names such as Expression, Japanese, Sentence, Meaning and English are also detected automatically.

For best results, format the English field like this:

First line: meaning of the complete sentence

Blank line

Remaining lines: vocabulary, grammar or scene notes

The notes are provided to the AI as context but are never read aloud directly.

SETTINGS

Tools ▸ SentenceSpeak ▸ Settings…

• AI: River API key, primary model, optional fallback model, Test API and Recent calls.
• Read Aloud: shortcut, Japanese/English field names, Japanese and Vietnamese voices, setup checks, self-test, Microsoft TTS updates and Clear Cache.
• Translate Popup: target language and keyboard shortcut.

Only one AI model is enabled by default. Adding a fallback model makes both models race for the fastest response and may charge both requests.

VOICES

SentenceSpeak uses Microsoft Edge TTS through the lightweight edge-tts client.

Default voices:

• Japanese: Nanami Neural at 1.0× speed
• Vietnamese: Nam Minh Neural at 1.2× speed

Optional voices include Keita for Japanese and Hoai My for Vietnamese.

No large voice model is downloaded. Internet access is required when generating new audio. Completed MP3 files are cached locally, so previously processed sentences can play immediately.

The audio cache is limited to approximately 300 MB. Older files are automatically removed when the limit is exceeded. You can remove all cached translations and audio manually with Settings ▸ Read Aloud ▸ Clear Cache.

Microsoft Edge TTS is an online service accessed through an unofficial client. Microsoft may change the service, although SentenceSpeak automatically checks for compatible client updates.

COST AND PRIVACY

• River AI is called only when K is pressed for a sentence that is not already cached.
• With the current prompt and DeepSeek pricing, a typical new card costs approximately $0.0001–$0.0002.
• Cached cards do not require another River request.
• Translate Popup and Explain requests are billed separately.
• The Japanese sentence, English meaning and relevant notes are sent to River for translation.
• Text that needs to be spoken is sent to Microsoft’s online TTS service to generate audio.
• No unrelated cards or collection data are sent.
• Your River API key remains in Anki’s local add-on configuration and is excluded from exported add-on packages.
• If River is unavailable or no API key is configured, card reading falls back to Google Translate. The result may be less natural and usually does not include explanation B.
• Translate Popup requires a River API key.

CACHE

Translations and generated audio persist across Anki restarts.

Cached content is automatically invalidated when relevant prompts, text, voices or speech speeds change. Use Clear Cache at any time if you want every card to be translated and generated again using the latest add-on version.

REQUIREMENTS

• Anki 25.07 or later.
• macOS, Windows or Linux.
• Internet access for new AI translations and uncached Microsoft speech.
• No separate Python, Azure account or offline voice model installation required.

The add-on has been tested on macOS and verified on Windows x64 with automated end-to-end tests covering installation, Microsoft voices, MPV playback, River translation, caching, chimes and the complete K-key reading sequence.

TROUBLESHOOTING

• Nothing happens after pressing K: open Settings ▸ Read Aloud ▸ Check Setup.
• Incorrect or outdated translation/audio: click Clear Cache and press K again.
• Microsoft speech stops working: click Update Microsoft TTS.
• To test every component, click Run Self-Test and send a screenshot of the result.
• Translation errors: check Settings ▸ AI ▸ Recent calls. Common examples include 401 for an invalid API key, 429 for a rate limit and 503 for temporary service unavailability.