Fine-tuning Whisper in more than one language

[andrespimartin]

Suppose I have a dataset in two or more languages (one of them under-represented in Whisper's pre-trained models), and I want to fine-tune those 2 or more languages to continue with a multilingual model and avoid catastrophic forgetting. Is fine-tuning possible?

Can I define the tokenizer and the processor without indicating the language?

[phineas-pta]

i don't think you can fine tune in one-click but you can do sequentially one language after another

[andrespimartin]

Hi @ILG2021

Sorry for the delay, I've had a busy week.
Your code seems correct to me. You don't need to do anything else: since the language label is added, the model will try to predict that label as well, so it will be computed in the loss calculation. I have used datasets with six languages with the samples randomly concatenated, and a batch size of 16 and it works fine.

Best!

[ILG2021]

the self.processor.tokenizer.pad in DataCollatorSpeechSeq2SeqWithPadding is language independent? and self.processor.tokenizer.batch_decode don't need to set_prefix_tokens?

[tuandattt]

Hi, I use @ILG2021 code and just change my dataset but i still get this error:
ValueError: Multiple languages detected when trying to predict the most likely target language for transcription. It is currently not supported to transcribe to different languages in a single batch. Please make sure to either force a single language by passing language='...' or make sure all input audio is of the same language.

How can I solve this?

[sproocht]

Hi @tuandattt ,
Try transformers version 4.37.2 (or earlier). I opened an issue about this limitation yesterday: huggingface/transformers#30654
All the best!

[sproocht]

Hi @tuandattt ,
This issue is now resolved in this fix and merged into the latest version of transformers.

[andrespimartin]

As long as your language is included in the Whisper language, it will be correctly encoded and decoded, so yes, it is language independent.

Regarding the self.processor.tokenizer.batch_decode, it is used when computing the metrics for the ASR task, so it is correct to skip special tokens (you only want to compute the metric of the ASR task).
Note that you are including the language label in the loss calculation. If you want to compute the metrics for the LID task, you will need to override or rewrite the compute metrics function.

[ILG2021]

hello @andrespimartin , I tried to train with language label, but the evaluate wer is not that stable. When I train without language column, the wer is reduce stablely. But when I use the language label, it may increase much. Have you ever encountered this problem?

[andrespimartin]

Hi @ILG2021, I have been checking my training logs and yes, it seems that when you add the language label the eval WER does not reduce in a stable way. I haven't had time to look at the code, but I understand that the language label affects the way the WER is calculated, because in tests the results are good.

Best

[ILG2021]

When I finetune multi language, I found that if the languages is more than 3, it is very easy to overfitting. Every language's data is about 1.5 hours to 3 hours. Anyone encounted this problem?

[pr0mila]

I’ve created a code-switched language dataset for fine-tuning Whisper, including audio data along with CSV and Parquet files, which I’ve stored on Hugging Face. After preparing the dataset, I fine-tuned the model for translation. You can explore the entire end-to-end project in my repo. Here’s the link to check it out: https://github.com/pr0mila/MediBeng-Whisper-Tiny