Welcome to the GitHub homepage for the highly optimized GGUF versions of AI4Bharat's flagship IndicConformer models for Hindi and Punjabi Automatic Speech Recognition (ASR).
Download the Models from Hugging Face Here
This model was explicitly converted and optimized to be run inside CAPIT—our custom desktop transcription software.
What is CAPIT? CAPIT is a fully offline, high-speed transcription application. It is built using Rust and Tauri, meaning it is incredibly lightweight, fast, and 100% private (no audio is ever sent to the cloud). Instead of writing python code or using terminal commands, you can simply load this model into the CAPIT desktop app and transcribe audio instantly with a beautiful user interface.
Check out the CAPIT Desktop App Here!
The models are converted directly from the official NeMo PyTorch checkpoints and are designed to be run locally with zero Python dependencies using the lightweight C++ C-API and the ggml execution provider.
Important
These GGUF models are fully metadata-driven. All configuration parameters, vocabulary strings, and featurizer window buffers are embedded directly into the GGUF binary, making it standalone.
indicconformer-hindi.gguf: Optimized for Hindi speech recognition.indicconformer-punjabi.gguf: Optimized for Punjabi speech recognition.
Evaluation results on standard test sets (Word Error Rate - WER & Character Error Rate - CER). Lower numbers represent better transcription accuracy.
| Model / Dataset | Hindi (Kathbath Test) | Punjabi (Kathbath Test) | Hindi (FLEURS Test) | Punjabi (FLEURS Test) |
|---|---|---|---|---|
| IndicConformer GGUF (WER) | 13.5% | 15.1% | 15.2% | 16.8% |
| IndicConformer GGUF (CER) | 5.2% | 6.8% | 5.9% | 7.4% |
Benchmarks were performed locally on a Windows workstation running an NVIDIA GeForce RTX 5070 Ti GPU and an Intel/AMD CPU.
- Test Audio Duration: 7.26 seconds (mono, 16000Hz).
- Metric: Real-Time Factor (RTF). Lower is faster.
| Configuration | Average Time (s) | Real-Time Factor (RTF) | VRAM / RAM footprint |
|---|---|---|---|
| Hindi (GPU / CUDA) | 0.602s | 0.083x (12x speed) | ~520 MB |
| Hindi (CPU) | 0.436s | 0.060x (16x speed) | ~140 MB |
| Punjabi (GPU / CUDA) | 0.654s | 0.090x (11x speed) | ~520 MB |
| Punjabi (CPU) | 0.415s | 0.057x (17x speed) | ~140 MB |
Note
For shorter clips (under 10 seconds), CPU is slightly faster because it avoids the initial GPU CUDA kernel compilation and memory-transfer latency. On longer files (e.g., 5+ minutes), GPU execution provides a massive speed improvement.
When deploying ASR systems locally, you must balance model size, system resources, accuracy, and execution latency.
| Feature / Model | IndicConformer GGUF (Ours) | OpenAI Whisper Large V3 | Gemma 4 Audio (12B) |
|---|---|---|---|
| Parameter Count | ~120M | ~1.5B | ~12B |
| RAM/VRAM Footprint | ~140 MB | ~3.1 GB | ~8.5 GB+ |
| Dependencies | None (Self-contained C++) | Python, PyTorch, Transformers | LLM Server (Ollama / HuggingFace) |
| Inference Mode | Real-time Streamable / Batch | Batch-only (seq2seq) | Batch-only (seq2seq) |
| Hindi Accuracies | Highly competitive local WER | ~11-12% WER (often struggles with local dialects) | High semantic accuracy, but prone to LLM paraphrasing |
Tip
Key Advantage: Our GGUF models run at over 15x real-time speed while consuming less than 5% of the memory footprint required by Whisper or Gemma 4, making them perfect for low-power edge devices and CPU-only systems.
To run these models, you need the parakeet-cli C++ execution engine.
- Go to the parakeet.cpp GitHub Repository.
- Follow their build instructions to compile the
parakeet-cliexecutable for your specific operating system (Windows/Linux/macOS). - Once compiled, open your terminal and run the models using the following commands:
parakeet-cli transcribe --model indicconformer-hindi.f32.gguf --input audio.wav --decoder ctc --lang hiparakeet-cli transcribe --model indicconformer-punjabi.f32.gguf --input audio.wav --decoder ctc --lang pa