bengali-whisper-medium-ct2

CTranslate2 int8 conversion of SayedShaun/bengali-whisper-medium, ready to use with faster-whisper — no conversion step needed.

Attribution. The underlying weights were trained by tugstugi (Erdene-Ochir Tuguldur), team Chimege — the 1st-place solution of the Bengali.AI Speech Recognition Kaggle competition. This repository is a format conversion only; see the base model for full attribution, citation, and details on how this checkpoint was built.

Usage

pip install faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel("SayedShaun/bengali-whisper-medium-ct2", device="cpu", compute_type="int8")
segments, _ = model.transcribe("clip.wav", language="bn", beam_size=1, vad_filter=True)
raw_text = " ".join(s.text for s in segments)

device="cuda" works too if you have a GPU. compute_type="int8_float16" on CUDA, "int8" on CPU.

This gives you a transcript without punctuation. For the full two-stage pipeline (ASR + punctuation restoration), see the base model's README — the punctuation stage is a separate, standalone package (asr-punct-restore) that works on this model's raw output the same way.

Verification

Transcribed against a real Bengali news clip and confirmed to match the transformers pipeline's output word-for-word.

How this repo was produced

Already done — nothing here to run yourself. Recorded for provenance, and for re-converting from a future revision of the base model:

pip install ctranslate2
ct2-transformers-converter --model SayedShaun/bengali-whisper-medium \
    --output_dir bengali-whisper-medium-ct2 --quantization int8 \
    --copy_files tokenizer.json preprocessor_config.json

License

Apache-2.0, following the base model.