W Whisperila AI v1.0.5 · macOS Download DMG
Startup Buildathon India 2026 · Working Native macOS Prototype

Voice AI for every Indian, not just English.

98% of India’s 958 million internet users prefer Indic-language content — yet AI tools still feel like English software wearing a robotic translated voice. Whisperila AI brings native हिन्दी (Hindi) dictation and a live autonomous computer-operating voice agent to everyone’s screen.

ARX Tech · Dual-Engine OS Architecture

"Why type or wrestle English menus? Just speak — and watch it teach, click, browse & build."

MODE 1 · ⌥ OPTION + SPACE
Whisperila Dictation & Meetings
41 ms p50 local CoreML (Parakeet Ultra, WhisperKit, Apple) + Gnani Prisma v3.0 / Sarvam AI with automatic local failover.
MODE 2 · FN HOTKEY
Whispy Autonomous AI Voice Agent
Gemini Live 2-way voice, Screen Vision + Apple Vision OCR, on-screen neon drawing tutor, Computer Use, Browser Agent & Live 3D CAD.
958M Active Indian internet users (IAMAI-Kantar ICUBE) — 57% live in rural Bharat.
98% Of Indian internet users consume & prefer Indic-language content over English.
$75.1B Combined 2030 TAM across Speech Dictation ($23.1B) + Voice AI Agents ($52B).
41 ms Measured p50 on-device recognition on Apple Silicon (Parakeet Ultra CoreML).
MODE 1 Parakeet Ultra CoreML MODE 1 WhisperKit Turbo INDIC Gnani Prisma v3.0 (indic-auto) INDIC Sarvam Saaras v2 MODE 2 Google Gemini 3.8 Live VISION ScreenCaptureKit + Apple Vision OCR DESKTOP Native macOS AX + CGEvent BROWSER CDP + DOM Label Automation 3D CAD AI 8-Primitive SceneKit Hologram AUDIO vDSP 512-pt FFT Human Barge-In 12+ INDIC हिन्दी · తెలుగు · தமிழ் · ಕನ್ನಡ · বাংলা · मराठी MCP Native JSON-RPC 2.0 Server MODE 1 Parakeet Ultra CoreML MODE 1 WhisperKit Turbo INDIC Gnani Prisma v3.0 (indic-auto) INDIC Sarvam Saaras v2 MODE 2 Google Gemini 3.8 Live VISION ScreenCaptureKit + Apple Vision OCR DESKTOP Native macOS AX + CGEvent BROWSER CDP + DOM Label Automation 3D CAD AI 8-Primitive SceneKit Hologram AUDIO vDSP 512-pt FFT Human Barge-In 12+ INDIC हिन्दी · తెలుగు · தமிழ் · ಕನ್ನಡ · বাংলা · मराठी MCP Native JSON-RPC 2.0 Server
01 · THE ACTUAL PROBLEM · REAL INDIA DATA

AI is exploding everywhere — yet 548M+ rural & vernacular Indians are locked out by English keyboards and robotic translated voices.

Earlier AI assistants were engineered for Silicon Valley English typists, then bolted onto rigid English-to-Indic TTS translators. They collapse on Hinglish, Tenglish, Tanglish, Indian accents, and lakhs/crores — and worst of all, they only output chat bubbles instead of actually operating your computer or teaching software step by step.

57%
Rural Bharat Majority

~548M of India’s 958M internet users are in rural India — growing 2×–4× faster than metros, yet underserved by desktop & agentic AI.

98%
Indic Content Preference

98% of Indian internet users consume Indic content (and 57% even in urban metros prefer regional languages over English — IAMAI/Kantar).

~10%
Fluent English Barrier

Only ~1 in 10 Indians is comfortable typing complex English prompts. Complex software menus (Blender, Excel, Figma, Gov portals) remain intimidating.

0%
Real Action in Chatbots

Robotic translated voice bots only talk — they cannot read your screen, draw an arrow over "Export", click forms, or model 3D objects live.

01 Indian Internet Users Consuming Indic Language Content (IAMAI-Kantar) 98.0%
02 Rural Share of India’s Active Internet Base (~548M Users) 57.0%
03 Urban Indians Preferring Regional / Indic Over English 57.0%
04 Share of AI Desktop Agents Built Primarily in English-Only UI/TTS ~92.0%
02 · THE SOLUTION · TWO ISOLATED MODES IN ONE NATIVE APP

Speak in your mother tongue.
Dictate instantly OR command an agent.

Click a hotkey card below to inspect live flow ↓
MODE 2 · FN HOTKEY
User Speaks Naturally

On-Device + Live Execution Pipeline

Real Result on Screen

            
03 · FULL STACK CAPABILITIES & MULTI-MODEL ARCHITECTURE

Every engine under one roof — built natively in Swift 6.

Click any capability card to inspect the exact models, APIs, and on-device algorithms running inside Whisperila AI.

Gemini Live 3.8 / 2.5 Native Audio

Natural Two-Way Indic Voice (Zero Robotic Translation)

Native audio-to-audio streaming in Hindi, Hinglish, Telugu, Tamil, Kannada, Malayalam, Marathi, Bengali, Gujarati, Punjabi, Urdu, Odia, and English. Sticky language & single-voice persona lock stops mid-session accent or language drift.

Architecture: Direct off-MainActor detachedReceiveLoop feeding a 200ms jitter-resilient 24 kHz WhispyPCM24RingBuffer via AVAudioSourceNode. CoreAudio auto-detects headphones vs built-in speakers; a 512-point Accelerate/vDSP FFT human voice classifier rejects fans, claps, and clicks while allowing crisp Live-DJ style human barge-in. Automatically pauses playing YouTube, Spotify, Apple Music, VLC, and QuickTime when you speak.
+ Inspect architecture
4-Stage Pixel Grounding

Step-by-Step Visual Tutor & Screen Drawing

Ask "how do I do X?" in any Indian language. Whispy reads your screen, navigates to the app/menu, draws exact neon boxes & arrows over real UI buttons, and auto-advances when you complete each step.

4-Stage Grounding: (1) macOS Accessibility @axN ref -> (2) Live AX text search -> (3) On-device Apple Vision OCR (VNRecognizeTextRequest sub-word boundingBox(for:) on full-res raw capture) -> (4) 0..1000 edge-ruler coordinates. Perceptual frameDelta auto-triggers [TUTOR_AUTO_CHECK].
+ Inspect architecture
Computer + Browser Use

Autonomous Desktop & Web Execution

Full click, double-click, right-click, drag, scroll, hover, shortcut chords, Unicode Indian-script typing, and DOM-level browser control (`click_text`, `fill_label`, tabs) gated by the Jev Safety Policy Router.

Architecture: Depth-20 BFS macOS AXUIElement tree snapshotter + CGEvent input engine + clipboard-preserving Unicode paste for Devanagari/Telugu/Tamil text + CDP agent-browser & native Safari/Chrome JavaScript fallback. Destructive shortcuts require voice confirmation.
+ Inspect architecture
Live 3D CAD Hologram

Voice-to-3D Modeling & Guided Inspection

Speak any real-world object ("build a two-stage rocket", "show a planetary gear or wind turbine"). Whispy composes accurate multi-part colored 3D models in a floating SceneKit hologram and highlights each part as it teaches.

Architecture: Gemini 2.5 Flash structured 3D Assembly JSON -> 8-primitive OBJ compiler (box, cylinder, cone, sphere, torus, prism, capsule, wedge) with full Euler rotation matrices, semantic PBR hex colors & part descriptions, plus Zoo.dev & cad CLI support.
+ Inspect architecture
Mode 1 · Local + Indic Cloud

Hybrid Dictation, Meetings & MCP Server

Sub-50ms local CoreML dictation (`Parakeet Ultra`, `WhisperKit Turbo`, `Apple Speech`) + `Gnani Prisma v3.0` & `Sarvam AI` with automatic local failover on cloud outages, SpeakerKit diarization, local `llama.cpp` notes, and JSON-RPC MCP server.

Architecture: Dual-stream mic + system audio capture via ScreenCaptureKit, Indian numbering formatter (`1 lakh`, `2,50,000`, `2.5 crore`), personal vocabulary CTC phrase boosting with 0.65 spelling guard, and 4 MCP tools for Claude Desktop & Cursor.
+ Inspect architecture
04 · MARKET CAP & ECOSYSTEM THESIS · CAPTURING BOTH MARKETS

Why we don’t fight in a single-feature dictation war: we capture BOTH Speech Dictation AND Agentic Voice AI.

Dictation apps only convert speech to text ($23.1B by 2030). Autonomous Voice AI Agents operate software and workflows ($52B by 2030 at 37%–46% CAGR). Whisperila AI is the OS-level Indic Voice Agent Ecosystem uniting daily fast typing (⌥ Space) with autonomous multi-modal action (Fn).

Market 1 · Dictation & Speech-to-Text
$5.5B → $23.1B (2030)

13.1%–19.1% CAGR (Grand View / MarketsandMarkets). High-frequency daily habit loop — every student, founder, doctor, and creator types 100+ times/day.

Market 2 · Autonomous Voice AI Agents
$3.4B → $52.0B (2030)

37%–46% CAGR (Grand View / MarketsandMarkets). Explosion of real-time voice-to-action computer, browser, tutoring, and enterprise workflow agents.

Whisperila AI · Combined Capture
$75.1B Combined 2030 TAM

Dictation is our zero-friction daily wedge; Whispy Agent + MCP plugins turn Whisperila into India’s open Voice Agent Ecosystem.

WHY WE ARE NOT JUST COMPETITORS — BUILDING THE VOICE AGENTS ECOSYSTEM

Other tools build closed voice silos. We are building the open Indic Voice Layer for every app on your OS.

We don’t compete with Blender, Excel, Chrome, VS Code, or India’s government/business portals — we make every one of them voice-operable and teachable in 12+ Indian languages. With built-in MCP (Model Context Protocol), bring-your-own Indic models (Gnani, Sarvam, Gemini Live, local Parakeet/Whisper), and visual overlay SDKs, developers and educators can plug domain skills directly into Whispy.

L1 OS Voice Wedge: Instant ⌥Space Indic Dictation & Meetings (Daily Retention) 100% Local/Hybrid
L2 Multimodal Action Runtime: Live Vision + OCR Drawings + Computer/Browser/CAD Fn Agent Mode
L3 Open Ecosystem: MCP Server + Pluggable Indic Models + Educator Skill Packs Platform Moat
05 · COMPETITIVE LANDSCAPE · APPLES-TO-APPLES

Why existing tools leave a massive white space.

Pure dictation apps stop at pasting text. Cloud chatbots can’t draw on your screen or operate desktop software. Whisperila AI bridges the entire stack with native Indic fluency.

Capability / Dimension Whisperila AI (Ours) Wispr Flow Superwhisper ChatGPT Voice / Gemini App Apple Siri
Mode 1: On-Device Private Dictation YES · 41ms CoreML Cloud-only Yes (Local) No global paste Limited
12+ Native Indian Languages + Code-Mixing YES · Sticky Lock Basic Hinglish Whisper only Voice/Lang drift Robotic / English-first
Mode 2: Live 2-Way AI Voice Agent + Barge-In YES · Fn Hotkey No No Voice chat only Turn-based
Live Screen Vision + On-Screen Drawing Tutor YES · AX + Vision OCR No No No screen drawing No
Computer Use + Browser Automation YES · Jev Policy Gate No No Browser cloud only Shortcuts only
Voice-to-3D CAD Live Hologram YES · Multi-part 3D No No No No
Cloud Outage Resilience (HTTP 500 Failover) Auto Local Failover Fails offline Local Fails offline Partial
Open Ecosystem (BYO Indic Keys + MCP Server) YES · Open Stack Closed SaaS ($15/mo) No MCP Closed SaaS ($20/mo) Closed OS
06 · VERIFIED PROTOTYPE BENCHMARKS · REAL HARDWARE DATA

Measured on Apple Silicon — zero cherry-picking.

Measured via WhisperilaCLI --bench across 355 LibriSpeech/MLS clips, 300 EdAcc accented conversational segments, and 597 held-out Indian English segments (12,757 scored words).

#1 Parakeet Ultra (CoreML Local) — LibriSpeech Clean WER (41 ms p50 · 150× realtime) 2.30% WER
#2 Parakeet Ultra (Held-Out Indian English) — 597 EdAcc segments (12,757 words) 11.51% WER
#3 Parakeet TDT v3 (Held-Out Indian English) — Baseline local model comparison 12.33% WER
#4 Vocabulary Phrase Booster (0.65 Floor + Spelling Guard) — Synthetic tech terms WER drop 14.29% → 6.44%

Measured directly on MacBook Air (Apple Silicon, macOS 26.5) across Whispy AI Agent subsystems in automated test suites (25+ passing tests).

A1 vDSP 512-pt FFT Human Voice Classifier — Rejects fans/claps/clicks per 32ms frame < 0.05 ms
A2 Human Barge-In Cutoff (Headphones / Speakers) — Smooth 8ms fade-out, 0% echo leak 96 ms / 160 ms
A3 On-Device Apple Vision OCR Target Grounding — Exact sub-word pixel bounding box ~45–90 ms
A4 Smart Change-Driven Vision Bandwidth Reduction — 800px delta-gated vs 2 FPS 1280px -90% Traffic
A5 Overlay CPU Usage (Dedicated 280×110 Cursor Window) — Reduced from 95.3% full-canvas spin < 1.0% CPU

Scored across 4 multi-speaker AMI meetings (5,917 words) and 6 two-person EdAcc video calls (5,008 words) with voice-change chunk splitting.

M1 EdAcc Video Calls — Speaker Label Accuracy (Split at voice changes) 98.1%
M2 EdAcc Video Calls — Speaker Purity / Attribution 93.8% / 93.3%
M3 AMI 4-Speaker Overlapping Meetings — Speaker Accuracy 95.4%
07 · THE REAL ENGINEERING STRUGGLE · BATTLE-TESTED ON REAL MACS

Building a real OS-level Voice Agent is brutally hard.
Here’s every wall we hit — and broke through.

Apple’s hardware VoiceProcessingIO fails with -10875 on Apple Silicon MacBook Air’s 7-channel internal mic array, stripping away OS echo cancellation and causing speaker bleed to trigger infinite 2-second interruption loops. Our fix: Decoupled CoreAudio AudioQueue (16 kHz / 32 ms) + off-MainActor AVAudioSourceNode 24 kHz ring buffer + CoreAudio headphone/speaker route detection + a 512-point vDSP FFT human-voice classifier + server-turn-aware zero-echo shielding.
Multilingual live models randomly flip into English when reading English UI buttons ("File", "Export") and shift vocal timbre when 2 FPS video streams force early contextWindowCompression. Our fix: Built WhispyLanguageTracker (Unicode Indic script + Hinglish + NLLanguageRecognizer), eliminated mid-stream clientContent desyncs, and introduced delta-gated 800px Smart Vision (-90% context bloat).
Raw vision-model coordinates drift by 40–120px on Retina displays and multi-monitor rigs. Our fix: 4-stage pixel grounding — depth-20 macOS Accessibility @axN refs -> live AX text lookup -> on-device Apple Vision OCR sub-word boundingBox(for:) -> edge-ruler normalized coordinates, paired with a tiny hardware-composited 280×110 cursor window (<1% CPU vs 95% full-screen redraw).
When third-party cloud speech endpoints (like legacy Gnani v2.5) fail with HTTP 500 or user Wi-Fi drops in rural/Tier-3 areas, cloud-only dictation apps die completely. Our fix: Upgraded Gnani to prisma-v3.0 + indic-auto, added transient 5xx retry, AND built automatic local failover to on-device Parakeet Ultra / Apple Speech so dictation never drops a single sentence.
BUILT BY ARX TECH · FOR EVERY INDIAN BUILDER, STUDENT & CREATOR

Voice is India’s native interface.
Whisperila AI makes it agentic.

Native macOS 14+ (Apple Silicon) · Apache-2.0 Open Architecture · Verified Whisperila-AI-1.0.5.dmg (22 MB)

Download Whisperila AI 1.0.5 (.dmg) Open 16-Slide Buildathon Pitch & Benchmarks