whisper-tamil-medium-ct2

CTranslate2 int8 conversion of vasista22/whisper-tamil-medium, ready for faster-whisper.

• Base model: vasista22/whisper-tamil-medium (Apache-2.0)
• Quantization: int8 (CPU-friendly; model.bin ≈ 774 MB)
• Tokenizer note: tokenizer.json here was generated with tokenizers 0.19.x so it loads on faster-whisper 1.0.x runtimes. (A tokenizer.json produced by newer tokenizers can raise data did not match any variant of untagged enum ModelWrapper on older runtimes.)

Usage

from faster_whisper import WhisperModel

model = WhisperModel("mohankumars26/whisper-tamil-medium-ct2", device="cpu", compute_type="int8")
segments, info = model.transcribe("audio.wav", language="ta", beam_size=1)
print("".join(s.text for s in segments))

Validation

On a Google FLEURS (ta_in) test clip this int8 model produced correct Tamil script — ~97% character match to the reference, avg_logprob ≈ −0.05.

Converted with:

ct2-transformers-converter --model vasista22/whisper-tamil-medium \
  --output_dir ct2-tamil --quantization int8 \
  --copy_files tokenizer.json preprocessor_config.json