Pandivo TV
Next-Gen AI-Powered Music & Continuous AI DJ Experience for Amazon Fire TV & Android TV

The Challenge
Living room television environments present unique engineering constraints: TV remotes suffer from low-gain Bluetooth microphones and 800ms-2500ms cloud STT lag; standard Whisper models produce phantom silence hallucinations on ambient noise; and complex 10-foot D-pad navigation easily suffers from focus traps and frame drops on low-power TV chipsets.
Engineering Solution
Engineered a native C++ JNI on-device speech engine (Whistle AI, ~16.9 MB) equipped with an 80Hz high-pass filter, energy-based VAD silence trimming, and peak AGC (0.85). Coupled with an autonomous Continuous AI DJ powered by AWS Bedrock (Nova Lite) and Gemini, a plug-and-play USB OTG music player, and local SQLite WAL storage synchronized non-blockingly to Amazon DynamoDB.
Key Product Capabilities
Continuous AI DJ & Smart Curation: Multi-stage pipeline powered by AWS Bedrock (Nova Lite) extracting mood, energy curves, and harmonic BPM transitions with dynamic radio-host commentary.
100% Offline Speech-to-Text: ~120ms Time to First Token with zero cloud latency and zero network dependencies via native C++ JNI Whistle engine (~16.9MB acoustic model).
Plug-and-Play USB Audio Player: Automatic OTG detection for USB thumb drives with offline ID3 metadata extraction for MP3, FLAC, WAV, and AAC.
Sub-Millisecond Hybrid Storage: Local SQLite in WAL mode ensures 60 FPS cold boots, mirrored asynchronously to Amazon DynamoDB cloud sync.
Physical TV Verified at 60 FPS: Tested and validated on physical Fire OS hardware (FOS 14 3P TV) with real Fire TV remotes in Amazon Appstore Quality Central.
Open Source Contribution: Extracted and published the standalone npm package react-native-whistle-tv-offline for the global TV developer ecosystem.
Code Implementation
// AudioPreprocessor.kt - Native DSP pipeline for living room TV remote microphones
package com.pandivo.whistle
class AudioPreprocessor {
companion object {
private const val HPF_ALPHA = 0.969f // 1-pole HPF (~80 Hz at 16kHz)
private const val NOMINAL_PEAK = 0.85f
private const val MAX_GAIN = 6.0f
private const val RMS_SILENCE_THRESHOLD = 0.010f
fun processAudio(rawPcm: ShortArray): PreprocessResult {
var prevInput = 0f
var prevOutput = 0f
var peak = 0f
var sumSquares = 0.0
// 1. 80Hz High-Pass Filter: Eliminate DC bias & HVAC low rumble
val filtered = FloatArray(rawPcm.size)
for (i in rawPcm.indices) {
val input = rawPcm[i] / 32768.0f
val output = HPF_ALPHA * (prevOutput + input - prevInput)
prevInput = input
prevOutput = output
filtered[i] = output
val abs = Math.abs(output)
if (abs > peak) peak = abs
sumSquares += (output * output).toDouble()
}
// 2. Dual-Stage Silence Gating: Reject silence hallucinations
val rms = Math.sqrt(sumSquares / rawPcm.size).toFloat()
if (rms < RMS_SILENCE_THRESHOLD && peak < 0.05f) {
return PreprocessResult(isSilent = true, audio = floatArrayOf())
}
// 3. Peak Automatic Gain Control (AGC) for distant TV remote mics
val gain = if (peak > 0.001f) Math.min(NOMINAL_PEAK / peak, MAX_GAIN) else 1.0f
for (i in filtered.indices) {
filtered[i] = Math.max(-1.0f, Math.min(1.0f, filtered[i] * gain))
}
return PreprocessResult(isSilent = false, audio = filtered)
}
}Technologies Used
Architecture Stack
60 FPS TV interface featuring spatial D-pad focus routing, zero focus traps, and interactive companion avatars.
1-pole 80Hz high-pass filter, VAD silence trimming, automatic gain control, and RMS silence gating.
Amazon Bedrock Nova Lite with Gemini 3.5 Flash fallback for vibe understanding and real-time DJ commentary.
Local-first instant query engine synchronized asynchronously to DynamoDB cloud tables.
