Publish Visar wake-word models: ONNX, TFLite, and fine-tuning state

.gitattributes

@@ -33,3 +33,37 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165137066.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165151990.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165156182.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165200415.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165205430.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165209065.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165214153.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165220006.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165228805.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165233919.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165240393.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165340217.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165344343.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165351382.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165357948.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165409386.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165419082.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165424062.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165429632.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165434870.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165440008.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165505293.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165523705.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165622234.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165627330.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165633167.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165645302.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165651345.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165655777.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165701273.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165705473.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/20260223_165717600.wav filter=lfs diff=lfs merge=lfs -text
+user_calibration_audio/normalized_voice_26.wav filter=lfs diff=lfs merge=lfs -text
+visar_edge.onnx.data filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,55 @@
+---
+language:
+- hi
+- en
+tags:
+- openwakeword
+- wake-word-detection
+- voice-assistant
+- edge-ai
+- onnx
+- tflite
+license: apache-2.0
+---
+
+# Visar (वी-सार) — Custom openWakeWord Edge Model
+
+A lightweight, high-accuracy wake-word detection model custom-trained for low-power edge SBCs, PCs, and offline home automation systems.
+
+## Model Highlights
+- **Target Phrase:** "Visar" / "वी-सार" / "Vee-saar"
+- **Acoustic Tuning:** Native Hindi cadence and Indian English phonetic variations (`hi-IN-Madhur`, `hi-IN-Swara`, `en-IN-Prabhat`, `en-IN-Neerja`).
+- **Ground Truth:** Real-world microphone samples convolved with MIT environmental impulse responses (RIRs).
+- **Hard Negatives:** Penalized against acoustically close Indian words (*vichar*, *vishal*, *vikas*, *bazaar*).
+- **Target Hardware:** Raspberry Pi, Linux Edge SBCs, Intel x86, Android, ARM64 microcontrollers.
+
+## Audio Input Requirements
+- **Sample Rate:** 16,000 Hz
+- **Channels:** 1 (Mono)
+- **Format:** 16-bit Signed Linear PCM
+
+## Repository Contents
+| File / Directory | Description |
+| :--- | :--- |
+| `visar_edge.onnx` | Production ONNX inference graph |
+| `visar_edge.onnx.data` | External weight buffer for ONNX runtime |
+| `visar_edge.tflite` | Quantized FlatBuffer model for mobile/embedded devices |
+| `training_config.yaml` | Exact hyperparameter configuration used during training |
+| `training_state/` | Step checkpoints and optimizer states for fine-tuning |
+| `user_calibration_audio/` | Microphone ground-truth calibration recordings |
+
+## Quickstart (Python Inference)
+```python
+import numpy as np
+import openwakeword
+from openwakeword.model import Model
+
+# Initialize wake word engine with custom Visar model
+oww_model = Model(wakeword_models=["visar_edge.onnx"])
+
+# Pass raw 16kHz 16-bit PCM audio frames (1280 samples / 80ms chunk)
+# audio_frame = np.frombuffer(mic_stream.read(1280), dtype=np.int16)
+# prediction = oww_model.predict(audio_frame)
+# if prediction["visar_edge"] > 0.5:
+#     print("Wake-word detected: Visar!")
+

training_config.yaml

@@ -0,0 +1,32 @@
+augmentation_batch_size: 16
+augmentation_rounds: 1
+background_paths:
+- ./audioset_16k
+- ./fma
+background_paths_duplication_rate:
+- 1
+batch_n_per_class:
+  ACAV100M_sample: 1024
+  adversarial_negative: 50
+  positive: 50
+custom_negative_phrases: []
+false_positive_validation_data_path: validation_set_features.npy
+feature_data_files:
+  ACAV100M_sample: openwakeword_features_ACAV100M_2000_hrs_16bit.npy
+layer_size: 32
+max_negative_weight: 3000
+model_name: visar_edge
+model_type: dnn
+n_samples: 5000
+n_samples_val: 500
+output_dir: ./my_custom_model
+piper_sample_generator_path: ./piper-sample-generator
+rir_paths:
+- ./mit_rirs
+steps: 35000
+target_accuracy: 0.5
+target_false_positives_per_hour: 0.2
+target_phrase:
+- vee_saar
+target_recall: 0.25
+tts_batch_size: 50

training_state/negative_features_test.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7af4ff28a9da26c2dd16c67b0050c4ff90814c298654c44013acdce1765a5411
+size 3072128

training_state/negative_features_train.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0134f18643998b0e0460605db43f4427ae8cb087b44c97b353ff892974d1c050
+size 30720128