Remove ASR pipeline, use model directly to set forced decoder ids

app.py

@@ -4,13 +4,18 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor
+from transformers import WhisperForConditionalGeneration, WhisperProcessor
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
+target_language = "french"
 # load speech translation checkpoint
-asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
+whisper_model_name = "openai/whisper-base"
+whisper_processor = WhisperProcessor.from_pretrained(whisper_model_name)
+whisper_model = WhisperForConditionalGeneration.from_pretrained(whisper_model_name)
+decoder_ids = whisper_processor.get_decoder_prompt_ids(language=target_language, task="transcribe")
 
 # load text-to-speech checkpoint and speaker embeddings
 processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
@@ -23,8 +28,10 @@ speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze
 
 
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "french"})
-    return outputs["text"]
+    input_features = whisper_processor(audio["array"], sampling_rate=16000, return_tensors="pt").input_features
+    predicted_ids = whisper_model.generate(input_features, forced_decoder_ids=decoder_ids)
+    translated_text = whisper_processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
+    return translated_text
 
 
 def synthesise(text):