Whisper Large-v3 / Turbo TR FT: Common Voice + Bilkent + Mozilla TR + Custom Corpus

# === Whisper Large-v3-Turbo TR FT — RTX 4090 ===
import torch
from transformers import (
    WhisperForConditionalGeneration,
    WhisperProcessor,
    Seq2SeqTrainingArguments,
    Seq2SeqTrainer,
    DataCollatorForSeq2Seq,
)
from datasets import load_dataset, Audio
 
# 1. Model + processor
model = WhisperForConditionalGeneration.from_pretrained(
    "openai/whisper-large-v3-turbo",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)
processor = WhisperProcessor.from_pretrained("openai/whisper-large-v3-turbo")
processor.tokenizer.set_prefix_tokens(language="turkish", task="transcribe")
 
# 2. Dataset — Common Voice TR
cv_tr = load_dataset("mozilla-foundation/common_voice_17_0", "tr", split="train+validation")
cv_tr = cv_tr.cast_column("audio", Audio(sampling_rate=16000))
 
# 3. Preprocess
def prepare(batch):
    audio = batch["audio"]
    inputs = processor.feature_extractor(audio["array"], sampling_rate=16000)
    batch["input_features"] = inputs.input_features[0]
    batch["labels"] = processor.tokenizer(batch["sentence"]).input_ids
    return batch
 
cv_tr = cv_tr.map(prepare, num_proc=8, remove_columns=cv_tr.column_names)
 
# 4. Training args
args = Seq2SeqTrainingArguments(
    output_dir="whisper-large-v3-turbo-tr",
    per_device_train_batch_size=8,
    gradient_accumulation_steps=2,
    learning_rate=1e-5,                      # Whisper FT için düşük lr
    warmup_steps=500,
    max_steps=5000,
    bf16=True,
    optim="adamw_torch",
    predict_with_generate=True,
    generation_max_length=225,
    save_steps=500, eval_steps=500,
    logging_steps=25,
    report_to="wandb",
)
 
trainer = Seq2SeqTrainer(
    model=model, args=args,
    train_dataset=cv_tr,
    data_collator=DataCollatorForSeq2Seq(processor),
    tokenizer=processor.feature_extractor,
)
trainer.train()
 
# Bench (RTX 4090 + Common Voice TR 180h):
# - Wall-clock: 18-22 saat
# - Peak GB: 16-18
# - WER pre-FT: 12.3%
# - WER post-FT: 5.8%