Update app.py

app.py

@@ -3,43 +3,51 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, WhisperForConditionalGeneration, WhisperFeatureExtractor, WhisperTokenizer, pipeline
+from transformers import (
+    SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, WhisperForConditionalGeneration, WhisperFeatureExtractor, WhisperTokenizer, pipeline,
+    BarkModel, BarkProcessor
+)
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
-# load speech translation checkpoint
-feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-base")
-tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-base", language="french", task="translate")
-
-model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
-forced_decoder_ids = tokenizer.get_decoder_prompt_ids(language="french", task="translate")
-asr_pipe = pipeline(
-    "automatic-speech-recognition", 
-    model=model, 
-    feature_extractor=feature_extractor,
-    tokenizer=tokenizer,
-    device=device
-)
 
 # load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("leo-kwan/speecht5_finetuned_voxpopuli_lt")
+# processor = SpeechT5Processor.from_pretrained("leo-kwan/speecht5_finetuned_voxpopuli_lt")
 
-model = SpeechT5ForTextToSpeech.from_pretrained("leo-kwan/speecht5_finetuned_voxpopuli_lt").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+# model = SpeechT5ForTextToSpeech.from_pretrained("leo-kwan/speecht5_finetuned_voxpopuli_lt").to(device)
+# vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+# embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+# speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
 
 def translate(audio):
+    # load speech translation checkpoint
+    feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-base")
+    tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-base", language="french", task="translate")
+    
+    model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
+    forced_decoder_ids = tokenizer.get_decoder_prompt_ids(language="french", task="translate")
+    asr_pipe = pipeline(
+        "automatic-speech-recognition", 
+        model=model, 
+        feature_extractor=feature_extractor,
+        tokenizer=tokenizer,
+        device=device
+    )
     outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate", "forced_decoder_ids": forced_decoder_ids})
     return outputs["text"]
 
 
 def synthesise(text):
-    inputs = processor(text=text, return_tensors="pt")
-    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+    model = BarkModel.from_pretrained("suno/bark-small")
+    processor = BarkProcessor.from_pretrained("suno/bark-small")
+    inputs = processor(text, voice_preset="v2/fr_speaker_1")
+    speech = model.generate(**inputs).cpu()
+    
+    # inputs = processor(text=text, return_tensors="pt")
+    # speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
     return speech.cpu()