What is the correct way to provide sequence bias to the pipeline for automatic-speech-recognitio task using Whisper Model

🤗Transformers
1

I want to use sequence_bias option which is available in generation config to bias few sequences during inference. I tried passing through the pipeline, doesn’t seems to work.Here is the sample code how i tried using sequence bias.

Details:
Transforms Commit: git+https://github.com/huggingface/transformers@1c7e5e236823cd38faac8115f96205a82c17fff9
Version: v4.31.0
Usage:
Pipeline Init:
generation_config: GenerationConfig = GenerationConfig.from_pretrained(
model_dir, config_file_name=“generation_config.json”,
max_new_tokens=255,
repetition_penalty=1)

speech_pipeline = pipeline(
“automatic-speech-recognition”,
model=model_dir, #point to whisper model dir locally
feature_extractor=feature_extractor,
tokenizer=self.tokenizer,
return_timestamps=False,
generate_kwargs={
“num_beams”: 5,
“generation_config”: self.generation_config
})

Inference:
speech_pipeline(samples,
max_new_tokens=max_new_tokens,
generate_kwargs={
“sequence_bias”: sequence_bias,
“num_beams”: 5,
“generation_config”: self.generation_config
})[“text”].strip()