Welcome to the GitHub homepage for the highly optimized GGUF (float32) conversion of the parthiv11/stt_hi_conformer_ctc_large_v2 model, a Conformer-CTC Automatic Speech Recognition (ASR) model specifically trained for Hindi and Hinglish (code-switched Hindi-English) speech.
Download the Model 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 model is designed for high-performance, local deployment with zero Python dependencies using the lightweight C++ parakeet-cpp engine (GGML backend).
Important
The GGUF file is fully standalone. Configuration parameters and the 128-token sentencepiece vocabulary are embedded directly inside the binary.
- Architecture: Conformer-CTC (Large)
- Parameter Count: ~120M
- Vocabulary Size: 128 tokens
- Training Data: ~2,900 hours of combined Hindi and Hinglish speech from the ULCA and Europal corpora.
- Input Format: 16,000 Hz mono-channel WAV audio.
These metrics are calculated using Greedy Decoding (without KenLM rescoring).
| Dataset | Word Error Rate (WER) | Character Error Rate (CER) |
|---|---|---|
| MUCS 2021 Blind Test | 9.37% | 2.74% |
| IITM 2020 Eval Set | 12.93% | 5.60% |
| Common Voice 6 Test | 13.16% | 4.50% |
| Common Voice 7 Test | 13.50% | 5.20% |
| Common Voice 8 Test | 14.37% | 5.95% |
Tested on a local edge device (NVIDIA RTX GPU & AMD/Intel CPU).
- Test Audio Duration: 7.26 seconds (mono, 16000Hz).
- Metric: Real-Time Factor (RTF). Lower is faster.
| Configuration | Average Inference Time | Real-Time Factor (RTF) | CPU / RAM Footprint |
|---|---|---|---|
| Hinglish CTC (GPU / CUDA) | 0.481s | 0.066x (15x speed) | ~520 MB VRAM |
| Hinglish CTC (CPU) | 0.383s | 0.053x (19x speed) | ~140 MB RAM |
Tip
The CTC architecture (Non-Autoregressive) is inherently much faster on CPU than traditional RNN-T or Seq2Seq decoders. For short clips, the CPU executes the model at nearly 20x real-time speed!
| Feature / Model | Hinglish Conformer-CTC GGUF (Ours) | OpenAI Whisper Large V3 |
|---|---|---|
| Parameter Count | ~120M | ~1.5B |
| RAM/VRAM Footprint | ~140 MB | ~3.1 GB |
| Code-Switching | Excellent native Hinglish support | Tends to force-translate into pure English or pure Hindi |
| Inference Mode | Fast Greedy CTC Decoding | Batch-only (seq2seq autoregressive) |
To run this model, 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 model using the following command:
parakeet-cli transcribe --model hinglish-conformer-ctc.f32.gguf --input audio.wav --decoder ctc --lang hi- Original Checkpoint: parthiv11/stt_hi_conformer_ctc_large_v2
- License: Under NVIDIA NGC Terms of Use / CC-BY-4.0.