Modelpublic

ThinkHome/whisper-large-v3-czech-cv13-ct2-fp16

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

1likes72downloads

Model Card

whisper-large-v3-czech-cv13-ct2-fp16

CTranslate2 / faster-whisper FP16 conversion of mikr/whisper-large-v3-czech-cv13.

The source model is a Czech fine-tuned Whisper Large v3 model based on openai/whisper-large-v3.

Source model card reports:

—Validation loss: 0.1283
—WER: 0.0789
—Tensor type: F16

Important: this repository contains a converted inference format. The source WER is copied from the original model card and should not be treated as a new benchmark of this converted repository unless independently evaluated on the same dataset.

Recommended use with faster-whisper

For best Czech transcription accuracy:

python

from faster_whisper import WhisperModel

model = WhisperModel(
    "ThinkHome/whisper-large-v3-czech-cv13-ct2-fp16",
    device="cuda",
    compute_type="float16",
)

segments, info = model.transcribe(
    "audio.aac",
    language="cs",
    task="transcribe",
    beam_size=5,
    temperature=0.0,
    vad_filter=True,
    condition_on_previous_text=True,
)

for segment in segments:
    print(segment.start, segment.end, segment.text)

For faster inference with a possible small accuracy tradeoff:

python

from faster_whisper import WhisperModel

model = WhisperModel(
    "ThinkHome/whisper-large-v3-czech-cv13-ct2-fp16",
    device="cuda",
    compute_type="int8_float16",
)

Recommended use with WhisperX

bash

whisperx audio.aac \
  --model "ThinkHome/whisper-large-v3-czech-cv13-ct2-fp16" \
  --language cs \
  --device cuda \
  --compute_type float16 \
  --batch_size 8 \
  --align_model comodoro/wav2vec2-xls-r-300m-cs-250 \
  --output_dir whisperx_out

With diarization:

bash

whisperx audio.aac \
  --model "ThinkHome/whisper-large-v3-czech-cv13-ct2-fp16" \
  --language cs \
  --device cuda \
  --compute_type float16 \
  --batch_size 8 \
  --align_model comodoro/wav2vec2-xls-r-300m-cs-250 \
  --diarize \
  --hf_token YOUR_HF_TOKEN \
  --output_dir whisperx_out

Precision note

This repository uses CT2 FP16. It is intended as the main quality-preserving deployment format for faster-whisper and WhisperX.

INT8 and INT8_FLOAT16 modes may be faster and smaller, but should be benchmarked separately if exact WER matters.

Recommended settings

Best WER / quality mode:

text

compute_type = float16
beam_size = 5
language = cs
temperature = 0.0
vad_filter = True
condition_on_previous_text = True

Fast mode:

text

compute_type = int8_float16
beam_size = 1 to 3
language = cs
temperature = 0.0
vad_filter = True
condition_on_previous_text = False