IndicWhisper GGML (whisper.cpp format)

This repository contains GGML model files for Vignesh-M/Indic-whisper.

Why `.bin` instead of `.gguf`?

While large language model runtimes (like llama.cpp) have migrated to the newer GGUF format, whisper.cpp historically and currently uses the GGML binary format specification for its speech-to-text model files.

Because convert-h5-to-ggml.py generates native GGML models rather than GGUF models, these files use the .bin extension (matching official whisper.cpp models like ggml-base.bin). This is the correct native format expected by your whisper-server sidecar.

Compatibility Note

The original PyTorch Transformers checkpoint from Hugging Face is not directly compatible with whisper.cpp or whisper-server. These files have been converted using whisper.cpp's convert-h5-to-ggml.py script into the bin format accepted by whisper-server builds.

Files Available

ggml-indic-whisper-medium-f16.bin: FP16 half-precision model (~1.5 GB). This model offers the highest accuracy matching the original fine-tune. Recommended for systems with modern GPUs (RTX 4000/5000) where native FP16 Tensor Cores are available.
ggml-indic-whisper-medium-q8_0.bin: Q8_0 (8-bit quantized) model (~823 MB). This model offers nearly identical accuracy (>99% of FP16) with half the memory footprint and faster execution on CPU.

Conversion & Quantization Process

The models were converted using an optimized version of the convert-h5-to-ggml.py script and quantized using the whisper-quantize tool:

python3 convert-h5-to-ggml.py [model_dir] [whisper_pkg_dir] [out_dir]
whisper-quantize ggml-indic-whisper-medium-f16.bin ggml-indic-whisper-medium-q8_0.bin q8_0

Supported Languages

Supports many of the required Indic languages through the Whisper multilingual tokenizer/model family, but language coverage and accuracy should be verified per target language.