droman42/microwakeword-irina-ru

microWakeWord — «Ирина» (Russian)

Quantized INT8 streaming TFLite wake-word model for the Russian «Ирина» (stress on the 2nd syllable: и-РИ-на), for ESP32-S3 / ESPHome and server-side pymicro-wakeword. First openly published Russian microWakeWord.

Files

• irina.tflite — streaming quantized model
• irina.json — v2 manifest (trained_languages: ["ru"], probability_cutoff 0.97, sliding_window_size 5)

Metrics (on real household recordings, not synthetic)

• Recall: 100% (16/16) real «Ирина» recordings at every cutoff through 0.97 (positive mean-prob 0.996) — including congested/child voices earlier models missed.
• FA/h on 3.13 h Russian ambient: 3.5 @0.90 · 1.9 @0.97 · ~1.0 @0.99 (thin sample — tune on device).
• Selected as the best of 15 random-subset runs on real recall (run 4; see repo).

Operating point — tune the cutoff on YOUR device

The manifest ships probability_cutoff = 0.97 — a deliberately recall-safe starting point. Real utterances score ~0.996 (tightly bunched just under 1.0), well clear of 0.97, so recall is 100% with comfortable margin. To trade a little of that margin for fewer false accepts, raise the cutoff toward ~0.99 on your own device as your real ambient dictates:

cutoff	recall (16 real)	FA/h*
0.97 (shipped)	100%	1.9
0.98	100%	1.3
0.99	100%	1.0
≥0.995	drops	—

*FA/h is from a thin 3.13 h ambient sample — indicative only; measure on your device. Above ~0.99 recall falls off a cliff, so don't exceed it. Secondary knob: raise sliding_window_size (5 → 7) for extra smoothing if FA is stubborn.

Why it's alive on real voices (and the French cautionary tale)

No open Russian microWakeWord existed; the closest attempt (French: 3 Piper voices and a widened SNR) trained cleanly yet was dead on device. This one is alive — and publishable — because of choices each checked by a native-Russian ear, not assumed:

• Correct stress, verified per engine. Wrong stress ⇒ mispronounced positives ⇒ dead model. Piper (espeak-ng) is fed plain Cyrillic; Silero is fed «Ир+ина» (+ before the stressed vowel, put_accent=False); Vosk's BERT accentizer handles it from plain text.
• A deliberately diverse, license-clean voice mix. Piper ruslan + Silero (xenia/kseniya/eugene) + a female-leaning Vosk-TTS minority (all Apache/CC-BY-NC-SA). The Vosk female voices restore the female-speaker diversity that a non-redistributable Piper voice would otherwise have provided — so the model generalizes to female speakers without any CC-BY-NC-ND source.
• Selected from a 15-run search, not one lucky run. Recall is stable run-to-run but false- accepts are a data-subset lottery; this artifact is the best of 15 random-subset runs, chosen on real recall/FA — not on the (misleading) synthetic eval.
• Curated hard-negatives — near-misses (Марина / Арина / Ирана / перина, diminutives, inflected forms) teach the model the boundary, not just the word.
• Battle-tested recipe, no heroics — SNR −5…+10 dB (not widened — widening is what broke the French run), negative class weight 20, 3 augmentation rounds, 12k steps.

Usage

from pymicro_wakeword import MicroWakeWord, MicroWakeWordFeatures
mww = MicroWakeWord.from_config("irina.json")   # feed 16 kHz mono 16-bit audio via MicroWakeWordFeatures

Or in ESPHome via the micro_wake_word component.

Training & license

Voice mix above (Piper ruslan + Silero + female Vosk-TTS minority) → CC BY-NC-SA 4.0; plus stock English negatives (kahrendt/microwakeword) and a Common Voice ru subset. Full recipe, per-run log, and methodology: https://github.com/locveil/wakeword-training (docs/recipe.md, EXPERIMENTS.md).