openai

/

whisper-large-v2

like 1.68k

Follow

OpenAI 4k

Hey @sboudouk ! We use the default generation kwargs in HF: https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationConfig.temperature

You can override these by passing them to the .generate method if you're using model/processor, or by forwarding generate_kwargs={"temperature": 1} to the pipeline if you're using the pipeline.

Do you have a code snippet for your comparison? If you could share it I'd be happy to provide some pointers as to where we can make changes!

Sure, thanks for the link , struggled to find the default generation kwargs, kind of new to HF.

Here is the snippet where I build my whisper model instance in my python code:

transcribe = model.transcribe(audio, language='fr', temperature=0.0)

So from my understanding, I need to add every correspounding kwargs as a parameter to transcribe just as I added the language and the temperature ?

Hey @sboudouk ! Yep, you can just pass generate_kwargs as required:

import torch
from transformers import pipeline
from datasets import load_dataset

device = "cuda:0" if torch.cuda.is_available() else "cpu"

pipe = pipeline(
  "automatic-speech-recognition",
  model="openai/whisper-large-v2",
  chunk_length_s=30,
  device=device,
)

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
sample = ds[0]["audio"]

generate_kwargs = {"language": "<|fr|>", "temperature": 0.0}  # add any other generate kwargs you require 
prediction = pipe(sample, generate_kwargs=generate_kwargs )["text"]

Can I use generate_kwargs in HF API?

Which API @Superchik ? For the pipeline, you can use the structure defined above. For feature_extractor + model, you can simply set language="fr" when you call model.generate:

model.generate(input_features, language="fr", task="transcribe")

See the following doc for more details: https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperForConditionalGeneration.forward.example

import requests

API_URL = "https://api-inference.huggingface.co/models/openai/whisper-large-v2"
headers = {"Authorization": "Bearer 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"}

def query(filename):
    with open(filename, "rb") as f:
        data = f.read()
    response = requests.post(API_URL, headers=headers, data=data)
    return response.json()

output = query("sample1.flac")