ADR 0002 System Tts

ADR-0002: Use the operating system's TTS, behind an interface

Source of truth: this wiki page.

• Status: Accepted
• Date: 2026-09-19

Context

Listening drills are a v1 feature, and Fluenough is language-agnostic. The TTS backend therefore has to cover as many languages as the deck format allows, which is to say: as many as possible.

Three options were evaluated.

Operating system TTS

android.speech.tts.TextToSpeech and AVSpeechSynthesizer, reached through the flutter_tts package (4.2.5, ~396k downloads, verified publisher — mature and widely used).

Adds nothing to the download size, requires no network, and has by far the widest language coverage available. On Android, voices are provided by Speech Services by Google and installed by the user through system settings; quality on modern devices is neural and good. Coverage and quality vary by device and by installed engine, which is the main drawback.

Kokoro-82M (neural, on-device)

An 82M-parameter Apache-2.0 model with unusually good quality for its size. A new release landed on 2026-04-23.

Two things rule it out for v1:

1. Coverage. Kokoro supports roughly ten languages — English (US/UK), Spanish, French, Italian, Portuguese (BR), Hindi, Japanese, Korean and Mandarin. Fluenough is language-agnostic. Shipping Kokoro as the primary engine would mean most decks have no audio at all.
2. Ecosystem maturity. Flutter bindings now exist — flutter_kokoro_tts and kokoro_tts_flutter — which is a genuine change from the situation a year ago, when this would have meant hand-rolled NDK and G2P work. But flutter_kokoro_tts is at version 0.0.1 with 81 downloads, requires espeak-ng data files to be configured manually, and pulls a ~50 MB model from Hugging Face on first run. That is not a dependency to build a core feature on yet.

Worth noting that Kokoro is not an alternative to espeak-ng so much as a consumer of it: espeak-ng acts as the grapheme-to-phoneme frontend inside the Kokoro pipeline. They sit at different layers.

Pre-recorded human audio

Best possible listening practice. Rejected for v1: it needs a network, per-clip licensing diligence, and media files that cannot live comfortably in git.

Decision

Use OS TTS as the sole v1 backend, behind a narrow TtsEngine interface (speak, stop, isLanguageAvailable, voicesFor).

Consequences

Listening drills ship in v1 for every language the user's device has a voice for, at zero download cost.

The interface is the important half of this decision: a KokoroTtsEngine can be added later as an optional download for the languages it covers, without touching drill code. Should the bindings mature, re-evaluate.

Audio quality is outside our control and varies by device. Decks may override per-card audio via the audio field for cases where this matters.

See also: Contributing · ADR-Index · Home