bengali-whisper-medium-ggml

whisper.cpp GGML weights converted from SayedShaun/bengali-whisper-medium, at five quantization levels. Each .bin is fully self-contained — mel filters, tokenizer vocabulary, and weights are all in the one file, so no other files from this repo or the original model are needed at inference time.

Not GGUF. These use whisper.cpp's own GGML binary format (magic ggml), not llama.cpp's GGUF container. They work with whisper-cli / whisper.cpp bindings, not with llama.cpp-family loaders.

Files

file	size	bits	notes
ggml-model-f16.bin	1.5 GB	16	unquantized baseline
ggml-model-q8_0.bin	786 MB	8	near-lossless
ggml-model-q6_k.bin	610 MB	6	k-quant; usually better quality-per-byte than q8_0/q4_0 at similar size
ggml-model-q4_0.bin	424 MB	4	good speed/size tradeoff for most CPU use
ggml-model-q2_k.bin	254 MB	2	aggressive — verify output quality on your own audio before relying on this one; low-bit quantization on a low-resource-language fine-tune can hallucinate or drop words rather than degrade gracefully

Benchmark on your target hardware before picking one — RTF (real-time factor) and accuracy both shift with CPU, thread count, and audio characteristics.

Usage

1. Get whisper-cli

git clone --depth 1 https://github.com/ggml-org/whisper.cpp
cmake -S whisper.cpp -B whisper.cpp/build -DCMAKE_BUILD_TYPE=Release
cmake --build whisper.cpp/build -j"$(nproc)" --target whisper-cli

2. Download a weight file

huggingface-cli download SayedShaun/bengali-whisper-medium-ggml \
  ggml-model-q4_0.bin --local-dir .

(swap ggml-model-q4_0.bin for whichever quant level you want)

3. Transcribe

./whisper.cpp/build/bin/whisper-cli \
  -m ggml-model-q4_0.bin \
  -f audio.wav \
  -l bn \
  -t 6            # thread count — set to your CPU's physical core count

Useful flags:

• -l bn — Bengali language hint (skip auto-detection)
• -t N — CPU threads to use
• -nt — suppress timestamps in output, plain text only
• -ng — force CPU even if built with GPU support (for comparison benchmarking)
• --output-srt — write subtitles alongside plain text

GPU inference

If whisper.cpp was built with -DGGML_CUDA=ON (NVIDIA) or -DGGML_VULKAN=ON (AMD/Intel/anything with a Vulkan driver), whisper-cli uses the GPU automatically — no extra flags needed. Pass -ng to force CPU-only for a comparison.

Python (via pywhispercpp or similar bindings)

from pywhispercpp.model import Model

model = Model("ggml-model-q4_0.bin", language="bn")
segments = model.transcribe("audio.wav")
for s in segments:
    print(s.text)

Provenance

Converted with whisper.cpp's models/convert-h5-to-ggml.py from the HF transformers-format checkpoint, then quantized with whisper-quantize. No retraining or fine-tuning was done in this conversion — these are the same weights as SayedShaun/bengali-whisper-medium, just reformatted and (for the quantized variants) compressed.

License

Apache-2.0, inherited from the base model.