Darek speaks with his own bundled voice, through our own Polish pronunciation rules

Context

Darek’s Replies were to be read out by the phone’s text-to-speech engine (ADR-0012). On most phones that engine is Google’s, and a phone without an offline Polish voice left Darek mute. That raised the question of whether a Command that calls may dial on a Reply nobody hears.

The owner decided on 2026-11-01 that Darek does not rely on Google, and that a user will be able to type their own Reply text. So fixed recorded clips are not enough: Darek needs a voice that reads any Polish text, inside the app, offline.

Read on the host on 2026-11-01:

• The voice. Piper’s pl_PL-mc_speech-medium (a VITS model in ONNX, 22.05 kHz), 63 MB. It renders 4.6 s of speech in 0.09 s on the PC CPU; the phone’s time is a phase 1 reading (R3).
• The pronunciation. Piper turns text into phonemes with eSpeak NG, which is GPL-3.0. Linking it would put Darek’s code under the GPL, which his all-rights-reserved licence cannot take. Polish spelling is regular enough for rules instead. A rule set written for this produces the model’s phoneme alphabet: letters and digraphs, soft i, nasal vowels by what follows them, voicing assimilation, final devoicing, penultimate stress and numbers as words. On the 3,000 most frequent Polish words it agrees with eSpeak exactly on 92.4%, with a phoneme error rate of 1.4%. Most differences are English names, abbreviations and eSpeak’s own slips (musimy is [muˈɕimɨ], not eSpeak’s [muˈsimɨ]). The rules were written from Polish orthography, not ported from eSpeak’s pl_rules or pl_list; eSpeak was only the comparison oracle on the host. They are committed as scripts/voice/g2p_reference.py.
• The runtime. onnxruntime-android from 1.29.0 on (still in 1.31.1) merges INTERNET, ACCESS_NETWORK_STATE and a ContentProvider, ai.onnxruntime.TelemetryInitializer, that starts at process start and posts usage data to Microsoft. 1.28.0 merges none of them. Its jni/arm64-v8a/libonnxruntime.so is 28.6 MB (x86_64: 34.6 MB), and every LOAD segment of both 64-bit libraries is aligned to 16 KB (0x4000).
• The sound. The owner heard the same sentence through our rules and through eSpeak, and went ahead with ours on 2026-11-01. On the phone the same day he compared the full model with a dynamic int8 quantisation of it (19 MB) and heard a big difference: the full model ships. On the test phone it is also the faster of the two (phase 1, R3).

Decision

• The Reply is spoken by Darek’s own voice: the mc_speech model as Piper publishes it, run by ONNX Runtime for Android (com.microsoft.onnxruntime:onnxruntime-android, MIT), and played through AudioTrack. Android’s TextToSpeech is not used. No other app is involved, and every phone has the voice.
• ONNX Runtime is pinned at 1.28.0, the last version without the telemetry provider. The manifest also removes ai.onnxruntime.TelemetryInitializer with tools:node="remove", beside INTERNET’s removal (ADR-0008), so a later version cannot switch it on silently. ACCESS_NETWORK_STATE stays Play Billing’s alone. A bump past 1.28.0 re-reads the AAR’s manifest first. 1.28.0’s 64-bit .so files are 16 KB page-aligned, which Play requires; a version change re-reads that too.
• Text becomes phonemes through our own rules, speak/PolishPhonemes.kt, mapped to the model’s ids from its phoneme_id_map: ^, then each codepoint’s id followed by the pad _, then $ (phase 1, section 3). No eSpeak, no GPL code.
• The model ships inside the app. It is fetched at build time from a release asset of this repo, pinned by SHA-256, as the recogniser’s model is (ADR-0009). It is loaded straight from the APK’s assets into the ONNX session, with nothing unpacked.
• Any Polish text can be spoken. Phase 3 lets the user write a Command’s Reply. The rules expand numbers, so a Reply with digits works.

Alternatives

The phone’s text-to-speech engine. It costs nothing in size and lets the user pick any engine, including free ones without Google (SherpaTTS, RHVoice, eSpeak NG from F-Droid). But most phones default to Google, the voice differs from phone to phone, and a phone without an offline Polish voice leaves Darek mute. Rejected on 2026-11-01.

Recorded clips of fixed Replies. Tiny and as clean as a recording can be, but they cannot say a Reply the user types.

Piper or sherpa-onnx with eSpeak NG. No rules to write, but GPL-3.0 inside the app.

A reduced-operator ONNX Runtime build (an ORT-format model and only the operators it uses). It brings the native library down to a few MB, at the cost of building ONNX Runtime ourselves: a custom native build in CI, redone for every version. Kept for the day the size reading hurts.

The model quantised to int8 (19 MB instead of 63 MB). The owner heard it as clearly worse on 2026-11-01, and on the test phone it loaded and spoke more slowly than the full model. Rejected: 44 MB is not worth a voice that sounds worse.

Other free Polish voices. Piper’s darkman and gosia carry the same licence question as mc_speech (below). Meta’s MMS-TTS and Coqui’s XTTS are non-commercial.

Consequences

• Daria’s voice, Piper’s gosia, is a second model of the same kind, run the same way (ADR-0016).
• The licence has a grey area, accepted knowingly. mc_speech’s recordings are CC0 and Piper’s voices are published under MIT. But the voice was fine-tuned from Piper’s English lessac voice, whose recordings (Blizzard 2013) are licensed for research only and exclude voice products. Whether that reaches a fine-tuned model is unsettled. The Licences screen credits the voice with both facts. The way out is a voice trained from scratch on the CC0 MC Speech recordings, which swaps a file and changes no code.
• About 92 MB more installed on an arm64 phone: the 63 MB model and the 28.6 MB libonnxruntime.so, both stored uncompressed in the APK. Phase 1 reads the real total (R10).
• ONNX Runtime’s version is held back on purpose. A newer one means reading its manifest for anything that reaches the network, the way ADR-0007 found datatransport behind billing.
• Synthesis runs on the phone’s CPU each time a Reply is spoken. Fixed Replies can be rendered once and cached in memory.
• The pronunciation rules are ours to maintain. A mispronounced word is a rule fix with a test, not a bug report to an engine.
• The no-voice case is gone. What remains is a voice that fails (the model does not load, or playback errors), which ADR-0011 treats like a Reply that finished at once.