ನುಡಿಯಕ್ಷರ — Kannada Speech-to-Text

ನುಡಿಯಕ್ಷರ (Nudiyakshara) converts spoken Kannada to Unicode Kannada text. Built by Sanchaya and Sanchi Foundation as a free, open-source tool for Kannada language preservation and research.

🌐 Live demo: dani.sanchaya.net

---

Features

Feature	How	Works on
Live mic recording	Browser Web Speech API (kn-IN)	Android Chrome, desktop Chrome/Safari
Audio file upload	Server-side OpenAI Whisper / mlx-whisper	Local server only
CLI batch transcription	transcribe.py	Any machine with Python

iOS Safari: Apple's Dictation service does not support Kannada. Use Android Chrome or desktop browsers for live recording.

---

Quick Start

Option 1 — GitHub Pages (live recording only, no install)

Open dani.sanchaya.net in Chrome or Safari. No installation needed.

Option 2 — Local server (live recording + file upload)

Requirements: Python 3.9+, Apple Silicon Mac recommended for file upload.

# Clone
git clone https://github.com/sanchaya/kn-voice-converter.git
cd kn-voice-converter

# Install dependencies
pip install -r requirements.txt

# Run (Apple Silicon — uses Metal GPU)
python app.py --backend mlx

# Run (CPU fallback)
python app.py --backend local

# Run on custom port
python app.py --port 8998

Open http://localhost:5000 in your browser.

Option 3 — CLI transcription

# Single file
python transcribe.py audio.mp3

# Batch folder
python transcribe.py ./recordings/ --batch --output ./transcripts/

# Choose model size
python transcribe.py audio.wav --model large-v3

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        Browser (dani.sanchaya.net)               │
│                                                                   │
│   Live recording  ──►  Web Speech API (kn-IN)  ──►  Transcript  │
│                         [browser-native, no server]              │
│                                                                   │
│   File upload     ──►  Flask POST /transcribe  ──►  Transcript  │
│                         [requires local server]                  │
└─────────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────────┐
│                     Local Server (app.py)                        │
│                                                                   │
│   GET  /           ──►  Main UI (build_html())                  │
│   GET  /about      ──►  about.html (Sanchaya info)              │
│   POST /transcribe ──►  do_transcribe() → JSON                  │
│   GET  /health     ──►  {"status": "ok"}                        │
└─────────────────────────────────────────────────────────────────┘

Transcription backends

The server supports three backends, set via BACKEND in app.py or --backend flag:

Backend	Flag	Best for	Notes
mlx	--backend mlx	Apple Silicon (M1–M4)	Uses Metal GPU + Neural Engine via mlx-whisper
api	--backend api	Cloud / accuracy	Requires OPENAI_API_KEY env var
local	--backend local	Any machine	CPU-only, slow on large files

Model sizes (mlx backend)

Model	Flag	Speed	Accuracy
turbo	--model turbo	Fast	Good (default)
large-v3	--model large-v3	Slow	Best
small	--model small	Fastest	Lower

---

Technology

• Web Speech API — browser-native speech recognition, lang="kn-IN", no server needed for live recording
• OpenAI Whisper — open-source (Apache 2.0) multilingual speech recognition model
• mlx-whisper — Apple MLX port, runs Whisper on Metal GPU and Neural Engine
• Flask + flask-sock — Python web server with WebSocket support
• Kannada prompt engineering — initial_prompt="ಕನ್ನಡ ಭಾಷೆಯಲ್ಲಿ ಮಾತನಾಡಿದ ಪಠ್ಯ:" biases Whisper toward Unicode Kannada output
• GitHub Pages — static hosting for the live demo; all assets (logo, icons) embedded as base64 data URIs in a single index.html

---

Project Structure

kannada-speech-to-text/
├── app.py              # Flask server — live UI + file upload endpoint
├── transcribe.py       # CLI tool — single file and batch transcription
├── index.html          # Static GitHub Pages app (live recording only)
├── about.html          # About page — Sanchaya mission, tech, support
├── assets.json         # Base64-encoded logos and icons
├── requirements.txt    # Python dependencies
├── CNAME               # dani.sanchaya.net (GitHub Pages custom domain)
└── push_to_github.sh   # Helper script to push to GitHub

---

Browser Compatibility

Browser	Live Recording	Notes
Android Chrome	✅	Best mobile experience
Desktop Chrome	✅	Full support including waveform
Desktop Safari	✅	Works on macOS
iOS Safari	❌	Kannada not in Apple Dictation language list
iOS Chrome	❌	Uses WebKit — same limitation as iOS Safari
Firefox	❌	Web Speech API not supported

---

Known Issues

• iOS Safari — service-not-allowed error because Apple's Dictation does not include Kannada. Cannot be fixed in software.
• Local server over HTTP on mobile — browsers block microphone on non-HTTPS. Use the GitHub Pages version (https://dani.sanchaya.net) instead.
• Firefox — Web Speech API is not implemented. File upload via local server works.

---

Supporting Sanchaya

ಸಂಚಯದ ಕೆಲಸಗಳಿಗೆ ಸಮುದಾಯದ ಕಾಣಿಕೆಗಳೇ ಬಲ.

File upload requires dedicated server hardware or GPU infrastructure — either a GPU cloud instance or a self-hosted machine. The compute cost is significant. Community funding will help Sanchaya invest in the right hardware and make this feature available to everyone for free.

👉 sanchaya.org/support-us

---

License

Code: MIT License
Content: Creative Commons Attribution 4.0

© 2026 Sanchaya & Sanchi Foundation