Apollo510B Voice-Activated Wearable with Offline Keyword Detection Development

Product concept and user experience

This case study illustrates the design method used for the compact pendant which is a voice-controlled AI device intended for the rapid capturing of authorized voice commands.

The pendant must be used by attaching it either to the user's garments or around the neck. While in the idle state, the gadget enters into a low power listening state and uses offline voice detection.

Once the pendant detects the required command and verifies it, it can give the user haptic or light feedback and starts the second step according to predetermined actions.

The product was designed so that the wake word itself could be processed locally. This reduces dependence on an internet connection and avoids continuously uploading ambient audio to a cloud service.

After activation, the user could choose between a short local recording, a BLE event sent to the mobile application or an explicitly enabled cloud-transcription workflow. The app showed clear microphone, recording and data-sharing states so that the user was not left guessing whether the device was listening or uploading audio.

Apollo510B hardware architecture

The Apollo510B became the choice for the primary processor for wearable technology due to its ideal mix of low-power embedded processing, audio connectivity, local AI capabilities, and BLE compatibility.

The reference model consisted of several microphones, a rechargeable lithium-polymer battery, a battery charger, a fuel gauge, flash memory, haptic motor driver, a status light-emitting diode/LED, and a BLE antenna network.

Multiple microphones were evaluated to improve audio capture under different wearing conditions. A pendant can be partly blocked by fabric, rotate on a lanyard or sit near a noisy jacket zip.

Hardware layout therefore considered microphone-port access, isolation from the haptic motor, vibration coupling, wind exposure and distance from the wearer's mouth. The design team also selected enclosure materials and acoustic membranes that balanced protection from dust and sweat with practical sound transmission.

The power architecture has been designed in a way that the requirements of always-on function are clearly separated from power demanding "active" recording behaviour. In the low-power mode, the device engaged only those hardware elements that were required for audio capture and keyword analysis.

As soon as the keyword has been recognized, the processing capacity could increase temporarily and initiate user feedback mechanism as well as establish or restore the BLE connection. Such state-oriented architecture becomes a while target of creating good battery life and responsiveness at the same time.

Offline keyword-detection pipeline

The keyword-detection workflow began with a carefully defined phrase. Short phrases that are too common create false activations, while long phrases add user friction.

The team developed a keyword that was distinctive and easy to pronounce. Subsequently, they created a plan that took into consideration the nuances of pronunciation of the product keyword, the volume of sounds produced, the distance from the microphone, and other parameters such as indoor noise, outdoors noise, music, TV speech, vehicle environments, and incidental conversations.

There were digital microphone streams that were captured with the help of audio interfaces and processed further using lightweight processing line. The scheme of firmware was capable of buffering the audio that had been captured.

The detector output a confidence value rather than a simple yes-or-no response. Application logic then used thresholding, consecutive-frame checks and cooldown rules before accepting the keyword.

This matters because accuracy alone is not enough. A wearable that activates from nearby conversation loses user trust. Conversely, a device that misses normal spoken commands becomes frustrating. We evaluated false accepts, false rejects, time to detection, wake-word detection while walking and behaviour in music or traffic noise.

Where microphone quality was inadequate, we investigated gain settings, enclosure changes and model retraining rather than hiding the issue in application logic.

Local inference, tuning and privacy

The local model has been evaluated with respect to its memory consumption, processing times and power usage on the target device.

The size of the model is inherently linked to the product decision made since a larger model might improve detection performance in tough conditions, but at the expense of higher energy consumption, longer startup times and increased development complexity.

In terms of the wearable device model, the objective was to control for a very specific inference task that could operate for long times without the requirement of full speech to text processing to take place on device.

The wake word fine-tuning involved real word recordings collected in a consent based testing procedure.

Evaluation was segmented by environment, accent group and distance, allowing the engineering team to identify where changes were needed. This method provides stronger evidence than tuning only with quiet office recordings.

Privacy was built into the interaction model. In standby, the device processed short audio windows locally for keyword detection rather than storing continuous ambient recordings.

When the wake word triggered a recording event, clear LED or haptic feedback communicated the active state. The mobile app allowed the user to disable voice activation, remove recordings and control whether any audio was transferred for cloud processing.

BLE, companion application and secure updates

BLE established the connection between the wearable technology and the user's smartphone. At the time of the setup, the app paired the gadget, performed the action, displayed the battery information and looked for any firmware updates. Whenever the device was in use, the wearable transmitted small messages, e.g. "keyword found" or "ready to record" instead of keeping a pointless continuous connection.

If recording upload was enabled, audio data could be transferred in chunks after user confirmation or when the app was available.

The design included transfer retries, sequence numbers and integrity checks so that a temporary Bluetooth interruption did not silently corrupt the user experience. Device time was synchronised from the phone to provide meaningful event timestamps.

A production-ready update system included signed image verification, version compatibility checks, battery safeguards before update, resumable transfer and recovery behaviour. Firmware-update security is especially important in a device with microphones because an unauthorised image must never be allowed to alter the stated privacy behaviour.

Validation and engineering results

Validation activity took into consideration power consumption while being in standby listening, keyword detecting, BLE advertising, connected BLE, audio recording and firmware update modes. The team performed acoustic testing in quiet rooms, cafés, outdoors, in vehicles and while wearing the device under various types of garments.

Also, we measured microphone obstruction, enclosure vibration, charging, BLE range and reconnection process.

The completed reference product demonstrated a practical pathway for an Apollo510B-based voice wearable: always-available offline keyword detection, explicit privacy behaviour, low-latency feedback and secure BLE integration.

The same platform can be adapted for a voice-note pendant, field-worker safety device, branded wellness accessory, accessibility controller or smart-home interaction wearable.