from typing import Dict, Any, List #from transformers import WhisperForCTC, WhisperTokenizer from transformers import WhisperForConditionalGeneration, AutoProcessor, WhisperTokenizer, WhisperProcessor, pipeline, WhisperFeatureExtractor import torch #from functools import partial #import torchaudio import soundfile as sf import io # Check for GPU #device = torch.device("cuda" if torch.cuda.is_available() else "cpu") class EndpointHandler: def __init__(self, path=""): self.tokenizer = WhisperTokenizer.from_pretrained('openai/whisper-large', language="korean", task='transcribe') self.model = WhisperForConditionalGeneration.from_pretrained(path) #self.tokenizer = WhisperTokenizer.from_pretrained(path) #self.processor = WhisperProcessor.from_pretrained(path, language="korean", task='transcribe') #self.processor = AutoProcessor.from_pretrained(path) #self.pipe = pipeline("automatic-speech-recognition", model=model, tokenizer=processor.feature_extractor, feature_extractor=processor.feature_extractor) self.feature_extractor = WhisperFeatureExtractor.from_pretrained('openai/whisper-large') # Move model to device # self.model.to(device) def __call__(self, data: Any) -> List[Dict[str, str]]: print('==========NEW PROCESS=========') #print(f"{data}") #inputs = data.pop("inputs", data) #print(f'1. inputs: {inputs}') inputs, _ = sf.read(io.BytesIO(data['inputs'])) #inputs, _ = sf.read(data['inputs']) #print(f'2. inputs: {inputs}') input_features = self.feature_extractor(inputs, sampling_rate=16000).input_features[0] #print(f'3. input_features: {input_features}') input_features_tensor = torch.tensor(input_features).unsqueeze(0) input_ids = self.model.generate(input_features_tensor) #(f'4. input_ids: {input_ids}') transcription = self.tokenizer.batch_decode(input_ids, skip_special_tokens=True)[0] #inputs, _ = torchaudio.load(inputs, normalize=True) #input_features = self.processor.feature_extractor(inputs, sampling_rate=16000).input_features[0] #input_ids = self.processor.tokenizer(input_features, return_tensors="pt").input_ids #generated_ids = self.model.generate(input_ids) # #transcription = self.pipe(inputs, generate_kwargs = {"task":"transcribe", "language":"<|ko|>"}) # #transcription = self.pipe(inputs) # #print(input) # inputs = self.processor(inputs, retun_tensors="pt") # #input_features = {key: value.to(device) for key, value in input_features.items()} # input_features = inputs.input_features # generated_ids = self.model.generate(input_features) # #generated_ids = self.model.generate(inputs=input_features) # #self.model.generate = partial(self.model.generate, language="korean", task="transcribe") # #generated_ids = self.model.generate(inputs = input_features) #transcription = self.processor.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] #transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0] return transcription #original __call__ # def __call__(self, data: Any) -> List[Dict[str, str]]: # inputs = data.pop("inputs", data) # # Preprocess the input audio # input_features = self.tokenizer(inputs, return_tensors="pt", padding="longest") # input_features = {key: value.to(device) for key, value in input_features.items()} # # Perform automatic speech recognition # with torch.no_grad(): # logits = self.model(**input_features).logits # predicted_ids = torch.argmax(logits, dim=-1) # transcription = self.tokenizer.batch_decode(predicted_ids)[0] # response = [{"task": "transcribe", "language": "korean", "transcription": transcription}] # return response