bluemorpholimited/localpolyglot-edge-models

🌐 LocalPolyglot

The World's First Fully Offline Multi-Language → Cantonese Real-Time Interpretation Engine

Engineered by Blue Morpho Limited

---

🦋 About Blue Morpho Limited

Blue Morpho Limited is a frontier-edge AI engineering company specializing in deploying production-grade neural inference systems on consumer mobile hardware. We believe the future of AI is local, private, and instant — no cloud, no latency, no compromise.

Our engineering philosophy:

Principle	Implementation
🔒 Absolute Privacy	Zero network permissions. Your voice never leaves the device.
⚡ Real-Time Performance	Sub-2-second end-to-end latency on mobile hardware.
🧠 Production-Grade AI	Fine-tuned + quantized models, not toy demos.
📱 Consumer-First Design	Runs on a single smartphone — no server, no GPU cloud.
🌐 Universal Accessibility	90+ source languages → native Cantonese colloquial output.

---

📱 The Product: LocalPolyglot APK

LocalPolyglot is a single-installation Android application that transforms any smartphone into a real-time multilingual interpreter — specifically optimized for translating foreign languages into authentic, colloquial Cantonese speech.

How It Works — Fully Automatic

The app automatically detects the language being spoken in the surrounding environment. No manual language selection is needed. The user simply turns on the app, and it:

  🎤 Microphone continuously captures ambient speech (ALL languages)
       ↓  (Whisper auto-detects which language is being spoken)
  🔇 VAD filters silence, detects end of speech
       ↓
  📝 ASR (Whisper-v3-Turbo) auto-detects language + transcribes speech to text
       ↓  (No user action needed — language is detected automatically)
  🔄 NLLB-200 LoRA translates detected language into Cantonese colloquial text
       ↓  (Translation is always to Cantonese yue_Hant)
  🔊 CosyVoice TTS synthesizes natural Cantonese speech (speed auto-matched to speaker)
       ↓  (Adaptive Speed: fast speaker → fast Cantonese, slow speaker → normal)
  🎧 Audio routed exclusively to wired/bluetooth headphones
       ↓  (Speaker is permanently locked — audio only goes to headphones)
  🔄 Microphone NEVER stops — continues listening while Cantonese plays

Key Differentiators

Feature	Cloud-Based Apps	LocalPolyglot
Network Required	Always	Never (after first launch)
Privacy	Voice sent to servers	100% on-device
Latency	3-10 seconds	< 2 seconds
Language Detection	Manual selection	Fully automatic (90+ languages)
Cantonese Quality	Written Chinese output	Authentic colloquial
TTS Speed Control	Fixed speed	0.5x – 2.0x adjustable + adaptive
Speaker Output	Default speaker	Headphone-only (speaker locked)
Simultaneous Listen+Speak	Sequential	Full-duplex (mic never stops)
Adaptive Speed Matching	Fixed speed	Auto-matches speaker pace
Cost	Monthly subscription	One-time install, zero recurring cost

---

📱 Minimum Phone Requirements

Before downloading, please check if your phone meets the minimum specifications below. LocalPolyglot runs on-device AI inference — it requires a powerful phone, not a cloud server.

Minimum Requirements (Will Run)

Component	Minimum	Why
OS	Android 13 (API 33) or newer	Requires foreground microphone service + modern audio APIs
CPU Architecture	ARM64 (arm64-v8a)	ONNX Runtime native libs are 64-bit only. 32-bit phones will NOT work.
RAM	8 GB	Models need ~4 GB at runtime. With < 8 GB RAM, the OS may kill the app (OOM).
Free Storage	10 GB	6 GB for model download + 4 GB overhead for app + cache.
SoC	Snapdragon 7 Gen 1 or equivalent	Real-time inference requires a mid-to-high-tier mobile SoC.
Headphones	Wired, Bluetooth, or USB required	The app does NOT output audio through the phone speaker. Headphones are mandatory.
Internet	Required on first launch only	For downloading ~6 GB of AI models from HuggingFace. After that, offline forever.

Recommended Requirements (Best Experience)

Component	Recommended	Why
OS	Android 14 (API 34)	Latest foreground service + audio routing improvements.
RAM	12 – 16 GB	Prevents OOM kills during long background sessions; allows larger model cache.
Free Storage	15 GB	Extra space for model updates + app cache.
SoC	Snapdragon 8+ Gen 1 or newer	Sub-2-second latency. Snapdragon 8 series recommended for real-time ASR + Translation + TTS.
Storage Type	UFS 3.1 or newer	Faster model loading (models are 6 GB; UFS 2.x will load slowly).
Headphones	Wired or high-quality Bluetooth (aptX/LDAC)	Lowest audio latency for Cantonese playback.
Battery	4000 mAh or larger	Continuous mic capture + AI inference drains battery. Larger battery recommended for all-day use.

Target Device (Development Reference)

LocalPolyglot was developed and tested on the following device:

Spec	Value
Device	iQOO Neo7 Racing Edition
SoC	Qualcomm Snapdragon 8+ Gen1
CPU	Kryo (1×3.2GHz + 3×2.75GHz + 4×2.0GHz)
GPU	Adreno 730
RAM	16 GB LPDDR5
Storage	256 GB UFS 3.1
Architecture	ARM64 (arm64-v8a)
OS	Android 13+ (OriginOS)
Compiler Flags	-O3 -ffast-math -funroll-loops -march=armv8.2a+dotprod

Phones That Will NOT Work

Phone Type	Reason
32-bit phones (armeabi-v7a only)	ONNX Runtime native libraries are 64-bit only.
Android 12 or below	Missing FOREGROUND_SERVICE_MICROPHONE type + modern audio routing APIs.
Phones with < 8 GB RAM	Insufficient memory for 3 concurrent AI models (~4 GB runtime).
iPhones / iPads	LocalPolyglot is Android-only. iOS does not support ONNX Runtime native inference.
Phones without headphone jack or Bluetooth	No way to output Cantonese audio (speaker is permanently locked).

Compatible Phone Examples

Phone	RAM	SoC	Compatible?
iQOO Neo7 Racing Edition	16 GB	Snapdragon 8+ Gen1	Optimal
Samsung Galaxy S23	8/12 GB	Snapdragon 8 Gen 2	Excellent
Samsung Galaxy S22	8 GB	Snapdragon 8 Gen 1	Good
OnePlus 11	16 GB	Snapdragon 8 Gen 2	Excellent
Google Pixel 7 Pro	12 GB	Tensor G2	Good
Xiaomi 13 Pro	12 GB	Snapdragon 8 Gen 2	Excellent
Samsung Galaxy A54	8 GB	Exynos 1380	Minimum (may be slow)
Older phones (< 8 GB RAM)	< 8 GB	Various	Not compatible

---

📥 Download & Installation Guide

Before downloading, please verify your phone meets the Minimum Phone Requirements above.

Step 1: Download the APK

File: LocalPolyglot-v1.0.0.apk
Size: ~19MB
URL: https://huggingface.co/bluemorpholimited/localpolyglot-edge-models/blob/main/LocalPolyglot-v1.0.0.apk

Note: The APK is small (19MB) because it does NOT include the AI models. Models are downloaded automatically on first launch (6GB).

Step 2: Enable Unknown Sources

Settings → Apps & Notifications → Special Access → Install Unknown Apps
→ Select your browser → Allow from this source
``

### Step 3: Install the APK

Method A: Direct install on phone

1. Open the downloaded APK file on your phone
2. Tap Install when prompted
3. Tap Open after installation completes

Method B: Install via ADB (from PC)

adb install LocalPolyglot-v1.0.0.apk ``

Step 4: Grant Permissions

Permission	Why It's Needed
Record Audio	Microphone for capturing speech
Notifications	Foreground service notification
Bluetooth Connect	Connect to Bluetooth headphones

Step 5: First Launch — Model Download

1. Open LocalPolyglot app
2. Tap the Start Interpretation button
3. App shows Downloading models... status
4. Models download automatically (~6GB, requires WiFi)
   - Whisper ASR: ~3.5GB
   - NLLB Translation: ~2.7GB
   - CosyVoice TTS: ~150MB
5. Download progress shown on screen (file by file)
6. After download: app works 100% OFFLINE forever
``

> Important: First launch requires internet connection for model download. After that, NO internet is ever needed again.

### Step 6: Connect Headphones

1. Connect wired, Bluetooth, or USB headphones to your phone
2. The app automatically detects headphone connection
3. Cantonese audio will ONLY play through headphones
4. The phone speaker is PERMANENTLY LOCKED
5. If no headphones connected, no audio will play (privacy protection) ``

---

📖 User Manual

Basic Operation

Step	Action	What Happens
1	Open LocalPolyglot app	App starts in standby mode
2	Tap Start button	App begins listening to surrounding speech
3	Speak any language near the phone	App auto-detects language and transcribes
4	Wait ~1-2 seconds	Cantonese translation appears on screen
5	Cantonese speech plays through headphones	Audio output at adaptive speed
6	Tap Stop button	App stops listening and exits

Understanding the Screen

UI Element	Meaning
Red button (Stop)	Stop interpretation
Green button (Start)	Start interpretation
Speed slider	Adjust TTS speed (0.5x – 2.0x), also used as baseline for adaptive speed
Source text	Shows what the speaker said (in original language)
Cantonese text	Shows the Cantonese translation
Latency display	Shows milliseconds from speech to Cantonese output
Status indicator	Shows current state: listening → translating → synthesizing → ready

Status Indicators

Status	Meaning
Standby	App is idle, not listening
Listening	Microphone is active, detecting speech
Translating	ASR + Translation in progress
Synthesizing	TTS generating Cantonese speech
Running	Pipeline active, ready for next utterance
Downloading	First launch: downloading models

TTS Speed Control

Speaker Speed	Adaptive Behavior	TTS Output
Fast speaker (5+ words/sec)	Speed multiplier ~2.0x	Fast Cantonese
Normal speaker (2.5 words/sec)	Speed multiplier ~1.0x	Normal Cantonese
Slow speaker (1.25 words/sec)	Speed multiplier ~0.5x	Slow Cantonese

Battery Optimization Setup

1. Tap Battery Optimization / Auto-Start Settings button in the app
2. Disable battery optimization for LocalPolyglot
3. Enable auto-start permission (iQOO/Vivo specific)
4. Lock the app in recent apps (swipe down on app card → lock icon)
``


---

## 🏗️ Technical Architecture

### Full-Duplex Multi-Threaded Pipeline with Adaptive Speed Matching

LocalPolyglot uses a **two-thread architecture** that enables true simultaneous listening and speaking. The microphone **never stops recording** while Cantonese TTS audio plays through headphones.

<div align="center">

```mermaid
graph LR
    subgraph Thread 1: Continuous Capture
    A[🎤 Microphone<br/>16kHz PCM] --> B[🔇 Silero VAD<br/>32ms frames]
    B -->|Speech end + duration| Q[📦 Utterance Queue<br/>Max 5 pending]
    end

    subgraph Thread 2: Inference Worker
    Q --> C[📝 Whisper-v3-Turbo<br/>Auto-detect language + transcribe]
    C --> S[🧮 Adaptive Speed<br/>words/sec → multiplier]
    S --> D[🔄 NLLB-200 LoRA<br/>→ Cantonese colloquial]
    D --> E[🔊 CosyVoice TTS<br/>ONNX VITS + Speed]
    E --> F[🎧 AudioTrack<br/>Headphone forced]
    end

    style A fill:#0ea5e9,color:#fff
    style B fill:#64748b,color:#fff
    style Q fill:#f59e0b,color:#fff
    style C fill:#9333ea,color:#fff
    style S fill:#ec4899,color:#fff
    style D fill:#3b82f6,color:#fff
    style E fill:#eab308,color:#fff
    style F fill:#22c55e,color:#fff

Automatic Language Detection

The app uses Whisper-v3-Turbo which has built-in automatic language detection. No manual language selection is needed:

Step	What Happens
1. Audio captured	Microphone records speech in any language
2. Whisper encoder	Computes mel-spectrogram from audio
3. Language detection	Whisper decoder auto-detects the spoken language (90+ languages)
4. Transcription	Speech is transcribed to text in the detected language
5. Language code	The detected language code is passed to NLLB
6. Translation	NLLB translates from detected language → Cantonese (yue_Hant)
7. TTS output	CosyVoice synthesizes Cantonese speech from translated text

---

🧠 Model Modules

1. ASR — asr_whisper_q4km/

Technical Details (click to expand)

Spec	Value
Base Model	openai/whisper-large-v3-turbo
Parameters	809M
Quantization	INT8 Dynamic
Language Detection	Automatic, 90+ languages
Quantized Size	3486.5MB
Estimated Latency	300-800ms

2. Translation — nllb200_lora_cantonese_colloquial/

Technical Details (click to expand)

Spec	Value
Base Model	facebook/nllb-200-distilled-600M
Fine-tuning	LoRA r=16, alpha=32
Target Language	yue_Hant (Cantonese)
Quantization	INT8 Dynamic
Quantized Size	2713.9MB

3. TTS — tts_cosyvoice_cantonese_onnx/

Technical Details (click to expand)

Spec	Value
Base Model	Sherpa-ONNX CosyVoice Cantonese
Adaptive Speed	Auto-matches speaker pace
Quantized Size	0.0MB

---

Performance Benchmarks

Pipeline Stage	Estimated Time	Technology
Audio Capture	< 100ms	OpenSL ES
VAD Detection	32-64ms	Silero VAD ONNX
ASR Recognition	300-800ms	Whisper-v3-Turbo ONNX INT8
Translation	200-450ms	NLLB-200 LoRA ONNX INT8
TTS Synthesis	100-300ms	CosyVoice ONNX INT8
Audio Output	< 50ms	AudioTrack
Total	700-1750ms	Target: < 2000ms

---

APK Specification

Property	Value
Package Name	com.localpolyglot
APK Size	~19MB
Model Download	~6GB on first launch, then offline forever
Architecture	arm64-v8a only
Min SDK	33 (Android 13)
Architecture	Full-Duplex + Adaptive Speed
Recurring Cost	Zero

---

Feature Matrix

#	Feature	Description
1	Continuous Ambient Listening	Microphone never stops
2	90+ Language Auto-Detection	Whisper auto-detects language
3	Cantonese Colloquial Translation	Authentic spoken Cantonese
4	Full-Duplex Pipeline	Mic keeps recording while TTS plays
5	Adaptive Speed Matching	TTS auto-matches speaker pace
6	Headphone-Only Output	Speaker permanently locked
7	Absolute Privacy	Zero network after model download
8	Boot Auto-Start	Auto-starts on device power-on
9	First-Launch Model Download	APK is small; models download from HuggingFace

---

Why LocalPolyglot Over Cloud Translation Apps?

Concern	Google Translate	ChatGPT Voice	LocalPolyglot
Requires Internet	Yes	Yes	No (after first launch)
Voice Privacy	Sent to Google	Sent to OpenAI	On-device only
Language Detection	Manual selection	Manual selection	Automatic (90+ languages)
Latency	3-10s	5-15s	< 2s
Recurring Cost	Free/Paid	$20/mo	Zero
Full-Duplex	No	No	Yes (mic never stops)
Adaptive Speed	No	No	Yes (auto-matches pace)

---

Ideal Use Cases

Scenario	How LocalPolyglot Helps
Tourist in Hong Kong	Speak any language, hear Cantonese through earphones
Medical Professional	Communicate with Cantonese-speaking patients offline
Shopping & Negotiation	Real-time price discussions without internet
Taxi & Transport	Tell the driver where to go in Cantonese
Restaurant Ordering	Ask about ingredients in authentic Cantonese
Business Meetings	Discreet real-time interpretation through earphones
Remote Areas	Full offline operation in tunnels, mountains, airplanes

---

Disclaimer & Limitation of Liability

IMPORTANT — PLEASE READ CAREFULLY

By downloading, installing, or using the LocalPolyglot application (the Software), you acknowledge and agree to the following terms. If you do not agree with these terms, do not install or use the Software.

1. Nature of the Software

LocalPolyglot is an AI-assisted real-time interpretation tool that utilizes machine learning models for automatic speech recognition, translation, and text-to-speech synthesis. The Software processes audio input and generates translated output automatically without human review or verification.

2. No Guarantee of Accuracy

• The translation output generated by the Software is produced by automated neural network models and may contain errors, inaccuracies, omissions, or mistranslations.
• The Software does not guarantee that any translation will be accurate, complete, reliable, or suitable for any particular purpose.
• Speech recognition, language detection, translation, and synthesis are inherently probabilistic processes and may produce incorrect results, especially with:
  • Heavy accents, dialects, or non-standard pronunciation
  • Background noise, overlapping speech, or poor audio quality
  • Rare or low-resource languages
  • Idiomatic expressions, slang, or culturally specific references
  • Technical, legal, medical, or other specialized terminology
  • Very short or very long utterances

3. Not a Substitute for Professional Interpretation

• The Software is not a certified or professional interpretation service.
• The Software must not be relied upon as the sole or primary means of communication in situations where accuracy is critical, including but not limited to:
  • Medical or healthcare settings (diagnosis, treatment, medication instructions, informed consent)
  • Legal proceedings (court hearings, contracts, legal advice, police interactions)
  • Financial transactions (investment decisions, banking, contract negotiations)
  • Emergency situations (calling emergency services, reporting crimes)
  • Safety-critical environments (aviation, maritime, construction, manufacturing)
• Users should always verify important translations with a qualified human interpreter or native speaker.

4. Limitation of Liability

To the maximum extent permitted by applicable law:

• Blue Morpho Limited (the Company), its directors, officers, employees, contractors, affiliates, and licensors shall not be liable for any:

  • Direct, indirect, incidental, special, consequential, or punitive damages
  • Loss of profits, revenue, business opportunities, or anticipated savings
  • Financial loss of any kind, including but not limited to losses arising from:
    • Misunderstandings caused by translation errors
    • Miscommunication during business, medical, legal, or financial interactions
    • Incorrect information relied upon by the user or any third party
    • Failure of the Software to function correctly, including crashes, freezes, or delayed output
    • Loss of data, model files, or device storage related to the Software
  • Personal injury or property damage arising from the use or inability to use the Software
  • Any claim by a third party arising from the user reliance on the Software output

• The total aggregate liability of the Company for any and all claims arising from the Software shall not exceed the amount paid by the user for the Software, which is zero (USD $0.00) given that the Software is provided free of charge.

5. No Warranty

• The Software is provided AS IS and AS AVAILABLE without warranties of any kind, whether express, implied, or statutory.
• The Company disclaims all warranties, including but not limited to:
  • Merchantability or fitness for a particular purpose
  • Accuracy, reliability, or completeness of translations
  • Non-infringement of third-party rights
  • Uninterrupted or error-free operation
  • Compatibility with any specific device or Android version beyond stated requirements

6. User Responsibility

• The user assumes full and exclusive responsibility for any decision, action, or inaction taken based on the Software output.
• The user is solely responsible for verifying the accuracy of any translation before relying on it.
• The user must not use the Software in situations where translation errors could result in harm, loss, or legal liability.

7. Third-Party Components

• The Software incorporates third-party open-source models and libraries (Whisper, NLLB-200, CosyVoice, ONNX Runtime, Silero VAD).
• The Company is not responsible for the accuracy, behavior, or limitations of these third-party components.
• Third-party model licenses apply separately to the model weights and are included in this repository.

8. Indemnification

• The user agrees to indemnify and hold harmless Blue Morpho Limited from any claims, damages, losses, or expenses (including legal fees) arising from:
  • The user use or misuse of the Software
  • The user reliance on translation output
  • The user violation of these terms

9. Acceptance of Terms

• By installing and using the Software, the user acknowledges that they have read, understood, and agree to be bound by this disclaimer.
• If the user does not agree, they must uninstall the Software immediately and cease all use.

This disclaimer is provided for informational purposes and does not constitute legal advice. The Company reserves the right to update this disclaimer at any time without prior notice.

---

License & Attribution

Component	License	Source
LocalPolyglot App Code	Apache 2.0	Blue Morpho Limited
Whisper-v3-Turbo	MIT	OpenAI
NLLB-200	MIT	Meta AI / Facebook
CosyVoice TTS	Apache 2.0	k2-fsa / Sherpa-ONNX
Silero VAD	MIT	Silero
ONNX Runtime	MIT	Microsoft

---

Blue Morpho Limited

Frontier-Edge AI — Local · Private · Instant

LocalPolyglot is a sophisticatedly engineered product by Blue Morpho Limited.

The microphone never stops. The translation never sleeps. The speed always matches. The privacy never compromises.

Last updated: 2026-09-16 22:58 UTC