Whisper Large v3 Turbo TR - Selim Çavaş

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.3123
• Wer: 18.9229

Intended uses & limitations

This model can be used in various application areas, including

• Transcription of Turkish language
• Voice commands
• Automatic subtitling for Turkish videos

How To Use

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "selimc/whisper-large-v3-turbo-turkish"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    chunk_length_s=30,
    batch_size=16,
    return_timestamps=True,
    torch_dtype=torch_dtype,
    device=device,
)

result = pipe("test.mp3")
print(result["text"])

Training

Due to colab GPU constraints I was able to train using only the 25% of the Turkish data available in the Common Voice 17.0 dataset. 😔

Got a GPU to spare? Let's collaborate and take this model to the next level! 🚀

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 4000
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.45.2
• Pytorch 2.4.1+cu121
• Datasets 3.0.1
• Tokenizers 0.20.1