Whisper Prompting

[sanchit-gandhi]

Feature request

Add prompting for the Whisper model to control the style/formatting of the generated text.

Motivation

During training, Whisper can be fed a "previous context window" to condition on longer passages of text.

The original OpenAI Whisper implementation provides the user with the option of passing an initial_prompt to the model. This prompt is replaces the "previous context window" during inference.

By passing the prompt as the "previous context window", the Whisper model conditions its generation on whatever text is passed as the prompt. This allows the user to control aspects of the generation, such as spellings of named entities and punctuation formatting (see openai/whisper#963 (comment)).

This is possibly a cheaper way of adapting the Whisper model to specific decoding constraints than fine-tuning.

This notebook demonstrates prompting with the initial codebase, and explains how this can be achieved for HF's Whisper: https://colab.research.google.com/drive/14FSeaoRvgs5arOTfiMQBnQ5NaLyma7Tq?usp=sharing

The proposed API for prompting would look something as follows:

1. Encode prompt text to prompt token ids (processor.get_prompt_ids) - this method is a wrapper around processor.tokenizer.__call__ that doesn't add the special token ids:

prompt = "IR, Newswire"
prompt_ids = processor.get_prompt_ids(prompt)

2. Pass the input audio and prompt token ids to the .generate method to get the predicted ids:

pred_ids = model.generate(input_features, prompt_ids=prompt_ids)

3. Decode the predicted ids and 'slice' off the prompt (we can do this by passing the prompt_ids):

pred_str = processor.batch_decode(pred_ids, prompt_ids=prompt_ids)

=> We would need to wrap all of this forced_decoder_ids logic into the generate method and update the processor/tokenizer accordingly.

Your contribution

Happy to guide the integration and review any PRs!

[sanchit-gandhi]

cc @hollance

[pmollerus23]

Hello, I'd like to pick up this issue!

[sanchit-gandhi]

Hey @mollerup23! Super cool! We would first need to update the generate modelling code to slide the forced decoder ids as explained in the notebook:

transformers/src/transformers/models/whisper/modeling_whisper.py

Line 1453 in d5de578

def generate(

And then add a new method in the tokenizer to ignore the prompt ids. Does this sound good to you?

[connor-henderson]

Hey @mollerup23 @sanchit-gandhi. Apologies, I'm not sure how picking these up works, I started working on it cause I saw there was no assignee and now have something I think is ready for review. Should I just keep it locally or push it up?

Totally fine with whatever, @mollerup23 commented first.

[pmollerus23]

@connor-henderson @sanchit-gandhi I have not yet started on this issue, feel free to push your commits and pick it up!

[pmollerus23]

I will continue to look into what @sanchit-gandhi mentioned in the meantime.

[connor-henderson]

Sounds good, thanks

[sanchit-gandhi]

Closed via #22496

[romitjain]

Hi @sanchit-gandhi and @connor-henderson
I saw the PR, but I was wondering if we also integrated always_use_initial_prompt and condition_on_previous_text to the API? If no, is there any active work going towards it?
Thanks

[sanchit-gandhi]

Hey @romitjain - we're working on integrating the OpenAI Whisper algorithm into Transformers, which will provide more support for these fine-grained decoding parameters! c.f. #27492

[M-Ali-ML]

Hey @romitjain - we're working on integrating the OpenAI Whisper algorithm into Transformers, which will provide more support for these fine-grained decoding parameters! c.f. #27492

are contribution allowed here? I'd like to help on that.