metadata
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)
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!")