tasbeeh-kws

Tasbeeh dhikr counting models (Arabic)

Two complementary on-device approaches to counting spoken dhikr (سبحان الله، الحمد لله، الله أكبر، أستغفر الله، لا إله إلا الله, plus three longer Azkar) over open-mic audio:

• dhikr_v2.onnx (recommended) — one utterance-level multi-class model (~1.5 MB): 9 classes (8 phrases + "none"). Handles any utterance length via a dynamic time axis — no fixed windows, no per-phrase heads.
• 8 v1 openWakeWord-style binary heads — fixed-window per-phrase classifiers (~860 KB each). Kept for always-on streaming use; weaker on long phrases.

v2: utterance-level multi-class model

Pipeline: Silero VAD (app-side) segments the mic stream → each speech segment is featurized with openwakeword's frozen speech_embedding into (T, 96) frames (0.125 s/frame) → small temporal Conv1d (kernel 5) → [mean, max] pooled 256-dim → MLP → softmax over 9 classes. Count = one increment per segment whose argmax is a phrase (with a ~1 s refractory and an undo button).

Hold-out results (synthetic, source-level split; see eval_report_v2.json):

Overall argmax accuracy 0.989; the only phrase→phrase confusions in the confusion matrix are 2 alhamdulillah + 1 allahuakbar → subhanallah; 12/1057 held-out negatives fire a phrase (1.1%). Guided mode (user picks one dhikr, app counts when its class prob ≥ 0.5) is directly read from this table.

Notably, the v2 design eliminated the v1 wide-window problem: the two long Azkar that scored ~0.55 as fixed-window heads reach 1.00 here, because a 7-second dua is just a longer (T, 96) sequence rather than a truncated window.

import onnxruntime as ort, numpy as np
sess = ort.InferenceSession("dhikr_v2.onnx", providers=["CPUExecutionProvider"])
# frames: (N, T, 96) raw openwakeword embedding frames for the whole utterance
probs = sess.run(None, {"input": frames.astype(np.float32)})[0]  # (N, 9) softmax

Training: same TTS positives as v1 (~5k augmented samples), ~10k Common Voice Arabic negatives as class "none", temporal Conv1d + pooled MLP (the fix for a pooled-features version that confused most short phrases into alhamdulillah — mean/max pooling over time destroys word order, and سبحان الله / الحمد لله / أستغفر الله all end in الله), collapse guard, 150 epochs AdamW.

v1 heads (always-on streaming alternative)

Rapid-tasbeeh detection (2 noisy simulated streams x 25 utterances, short heads): subhanallah 34/50, alhamdulillah 20/50, allahuakbar 24/50, la_ilaha_illa_allah 40/50, astaghfirullah 36/50.

Overlap note: the long dhikr لا إله إلا الله وحده لا شريك له… contains لا إله إلا الله as its opening — when the 56-frame head fires, count that dhikr and suppress the 16-frame head for that burst.

Real-audio evaluation harness

• mZahran001/tasbeeh-kws-recorder — browser recorder + eval tabs that score any recording (or the whole collected set) against all 8 v1 heads and (guided mode) any class of v2.
• mZahran001/tasbeeh-kws-real — recordings collected so far. All numbers above are synthetic; before production, record 50–100 real utterances per phrase on target phones via the recorder Space and re-tune thresholds.

Training data and recipe

• Positives: 449 + 252 synthetic clips (SWivid/Habibi-TTS, F5-TTS; 7 Arabic dialects, 4 speeds, 3 text variants / 3 seeds), RIR-reverberated and noise-augmented. Dataset: mZahran001/tasbeeh-kws-data (includes scripts/train_v2.py).
• Negatives: ~10k 2–4 s Common Voice Arabic segments (validation split) + openWakeWord FP-validation features (v1).
• v1 training: openWakeWord Model.auto_train, 25k steps/phrase, with a collapse guard.
• v2 training: scripts/train_v2.py — source-level hold-out split, audiomentations augmentation, frame-level standardization baked into the ONNX graph, dynamic-T export verified vs PyTorch.

Caveats

• All evaluation is synthetic. The recorder Space is the gate to real numbers.
• v2's guided FP rates are computed against synthetic negatives only.
• Both models assume openwakeword's speech_embedding feature pipeline (openwakeword >= 0.6 from GitHub main).

Files

• dhikr_v2.onnx — v2 utterance-level model (input: (N, T, 96) float32, any T; output: (N, 9) softmax, classes = 8 phrases + "none")
• eval_report_v2.json — v2 hold-out metrics
• tasbeeh_<phrase>.onnx — 5 short v1 heads (input: (N, 16, 96) float32)
• dhikr_<phrase>_<frames>.onnx — 3 wide-window v1 heads (input: (N, 12|56|32, 96) float32)
• eval_report.json, eval_report_new.json — v1 eval reports