whisper-telugu-medium-ct2

A CTranslate2 conversion of vasista22/whisper-telugu-medium, for use with faster-whisper.

The weights are unchanged. Only the format differs: CTranslate2 is what faster-whisper loads, and converting once means every machine that needs the model can fetch it ready to run instead of converting it again.

Usage

from faster_whisper import WhisperModel

model = WhisperModel("kavithajohn/whisper-telugu-medium-ct2",
                     device="cuda", compute_type="float16")
segments, info = model.transcribe("audio.wav", language="te")
for segment in segments:
    print(segment.text)

Conversion

python -m ctranslate2.converters.transformers \
  --model vasista22/whisper-telugu-medium \
  --output_dir whisper-telugu-medium-ct2 \
  --quantization float16

float16 because that is what the model is loaded with on any GPU of 6 GB or more. On a small card, load it with compute_type="int8_float16" instead — the same files serve both.

A note on the tokenizer

The source repository publishes weights without a tokenizer.json, which is common and perfectly valid, but faster-whisper needs one beside the weights. The tokenizer.json here is taken unmodified from openai/whisper-medium — the model this was fine-tuned from, and whose tokenizer is identical across a Whisper generation. preprocessor_config.json comes from the source repository.

Licence and attribution

Apache-2.0, inherited from the source model. All credit for the fine-tune itself belongs to vasista22; this repository contributes nothing but a format conversion.