ವಾಣಿ · Kannada FastConformer ASR | ARTPARK IISc & OpenVoiceOS

WebML-Kit On-Device Engine (WASM SIMD / WebGPU)

OpenVoiceOS/artpark-iisc-vaani-fastconformer-kn-onnx · 100% Client-Side

⚡ Detecting Hardware... Checking Cache...

Zero cloud calls. Audio stays strictly local in your browser. Model weights (~476 MB INT8) are streamed directly from Hugging Face and saved in the browser's Cache API for instant subsequent sessions.

Acoustic Input Studio

16kHz Mono PCM

Ready

00:00.0

Project VAANI Verified Dialects 4 Samples

Urban · Bengaluru ಬೆಂಗಳೂರು ಕರ್ನಾಟಕದ ರಾಜಧಾನಿ "Bengaluru is the capital of Karnataka"

Science · Technical ವಿಜ್ಞಾನ ಮತ್ತು ತಂತ್ರಜ್ಞಾನ ನಮ್ಮ ಭವಿಷ್ಯ "Science & technology shape our future"

Conversational ನಾಳೆ ಬೆಳಗ್ಗೆ ಹತ್ತು ಗಂಟೆಗೆ ಸಭೆ ನಡೆಯಲಿದೆ "The meeting will occur tomorrow at 10 AM"

Agriculture · Malnad ಈ ವರ್ಷ ಮುಂಗಾರು ಮಳೆ ಉತ್ತಮವಾಗಿ ಸುರಿದಿದೆ "This year monsoon rains were abundant"

Transcribed Kannada (ಲಿಪ್ಯಂತರಣ)

BPE Tokenized

ಬೆಂಗಳೂರು ಕರ್ನಾಟಕದ ರಾಜಧಾನಿ

"Bengaluru is the capital of Karnataka"

Kannada Akshara & Subword Decomposition

▁ಬೆಂ #1024 ಗಳೂ #2491 ರು #6 ▁ಕರ್ನಾ #891 ಟಕದ #421 ▁ರಾಜ #310 ಧಾನಿ #772

Inference Time

142 ms

Audio Length

2.40 s

RTF (Real-Time Factor)

0.059x

ONNX Inference Pipeline

NeMo FastConformer-TDT Architecture Flow

Transducer with Duration Prediction

Stage 01

Audio Preprocessing

16kHz audio input converted to 128-channel Mel filterbanks using 25ms windows with 10ms frame stride.

128 Mel bins · 16,000 Hz

Stage 02 · INT8

FastConformer Encoder

Depthwise separable convolutions + relative multi-head self-attention with 8x subsampling.

encoder-model.int8.onnx (476 MB)

Stage 03 · TDT

Decoder Joint Network

Joint network emits both token probabilities and duration predictions, skipping blank frames in O(1).

decoder_joint-model.int8.onnx (5.4 MB)

Stage 04

BPE Detokenization

Subword tokens mapped through SentencePiece vocabulary back to Unicode Kannada text strings.

vocab.txt · Max 10 tokens/step

Deployment & SDK

Zero-Friction Inference Snippets

Python · OpenVoiceOS · cURL

# 1. Install official onnx-asr engine
pip install onnx-asr

# 2. Recognize Kannada speech in 3 lines of Python
import onnx_asr

model = onnx_asr.load_model("OpenVoiceOS/artpark-iisc-vaani-fastconformer-kn-onnx")
result = model.recognize("kannada_audio.wav")

print(f"Kannada: {result}")
# Output: "ಬೆಂಗಳೂರು ಕರ್ನಾಟಕದ ರಾಜಧಾನಿ"
      

# Install OpenVoiceOS speech-to-text plugin
pip install ovos-stt-plugin-onnx-asr

# In your mycroft.conf / ovos.conf:
{
  "stt": {
    "module": "ovos-stt-plugin-onnx-asr",
    "ovos-stt-plugin-onnx-asr": {
      "model": "OpenVoiceOS/artpark-iisc-vaani-fastconformer-kn-onnx",
      "lang": "kn",
      "quantized": true
    }
  }
}
      

// 100% Client-Side Kannada ASR with webml-kit and ONNX Runtime Web
import { createOnnxPipeline } from './webml-kit.js';

const pipeline = await createOnnxPipeline({
  task: 'automatic-speech-recognition',
  modelId: 'OpenVoiceOS/artpark-iisc-vaani-fastconformer-kn-onnx',
  modelFile: 'encoder-model.int8.onnx',
  vocabFile: 'vocab.txt',
  cache: true, // Cached into browser Cache API automatically
  device: 'wasm' // or 'webgpu'
}, (progress) => {
  console.log(`Downloading ${progress.file}: ${progress.percent}%`);
});

// Feed 16kHz Float32 PCM directly from Web Audio API microphone / file
const result = await pipeline(pcmFloat32Array);
console.log('Recognized Kannada:', result.text);
      

# 1. Download sample Kannada 16kHz audio directly from this Space
curl -O https://gnumanth-artpark-iisc-vaani-fastconformer-kn-onnx.static.hf.space/kannada_sample.wav

# 2. Run local on-device inference with onnx-asr CLI / Python
python3 -c "import onnx_asr; m = onnx_asr.load_model('OpenVoiceOS/artpark-iisc-vaani-fastconformer-kn-onnx'); print(m.recognize('kannada_sample.wav'))"

# Note: This is an on-device ONNX model running in-browser via WebML-Kit.
# Static HF Spaces do not host a POST endpoint (POST returns 404/405).
      

ವಾಣಿ (Vaani) FastConformer

ಕನ್ನಡ ವಾಕ್-ಗುರುತಿಸುವಿಕೆ FastConformer Acoustic Model

Acoustic Input Studio

Transcribed Kannada (ಲಿಪ್ಯಂತರಣ)

NeMo FastConformer-TDT Architecture Flow

Zero-Friction Inference Snippets