[whisper] fix multilingual fine-tuning

[sanchit-gandhi]

What does this PR do?

Fixes #30654.

The script previously failed with the following args, but passes after this fix:

python run_speech_recognition_seq2seq.py \
	--model_name_or_path="openai/whisper-small" \
	--dataset_name="mozilla-foundation/common_voice_11_0" \
	--dataset_config_name="hi" \
	--language="hindi" \
	--task="transcribe" \
	--train_split_name="train+validation" \
	--eval_split_name="test" \
	--max_steps="5" \
	--output_dir="./whisper-small-hi" \
	--per_device_train_batch_size="16" \
	--gradient_accumulation_steps="2" \
	--per_device_eval_batch_size="16" \
	--logging_steps="25" \
	--learning_rate="1e-5" \
	--warmup_steps="2" \
	--eval_strategy="steps" \
	--eval_steps="5" \
	--save_strategy="steps" \
	--save_steps="5" \
	--generation_max_length="20" \
	--preprocessing_num_workers="16" \
	--max_duration_in_seconds="30" \
	--text_column_name="sentence" \
	--freeze_feature_encoder="False" \
	--gradient_checkpointing \
	--fp16 \
	--overwrite_output_dir \
	--do_train \
	--do_eval \
	--predict_with_generate \
	--use_auth_token

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

Thanks for fixing!

[sproocht]

Thank you @sanchit-gandhi for fixing the issue and @amyeroberts for approving the changes!
My latest evaluation tests are positive. Well done!