Update app.py

app.py

@@ -2,8 +2,7 @@ import gradio as gr
 import numpy as np
 import torch
 from datasets import load_dataset
-
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+from transformers import AutoProcessor, AutoModel, pipeline, MarianMTModel, MarianMTTokenizer
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
@@ -12,38 +11,44 @@ device = "cuda:0" if torch.cuda.is_available() else "cpu"
 asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
 
 # load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+processor = AutoProcessor.from_pretrained("suno/bark-small").to(device)
+model = AutoModel.from_pretrained("suno/bark-small").to(device)
+
+# load MartianMT model for translating English to Hindi.
+martian_mt_model = MarianMTModel.from_pretrained("AbhirupGhosh/opus-mt-finetuned-en-hi")
+martian_mt_tokenizer = MarianTokenizer.from_pretrained("AbhirupGhosh/opus-mt-finetuned-en-hi")
+
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+def translate_english_to_hindi(english_text):
+    tokenized_text = martian_mt_tokenizer.encode(english_text, return_tensors="pt")
+    generated_token_ids = martian_mt_model.generate(tokenized_text)
+    hindi_text = martian_mt_tokenizer.decode(generated_token_ids.numpy()[0])
+    hindi_text = hindi_text.replace("</s>", "")
+    hindi_text = hindi_text.replace("<pad>", "")
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+    return hindi_text
 
 
-def translate(audio):
+def translate_to_english(audio):
     outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe"})
     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)
-    return speech.cpu()
-
+    speech_values = model.generate(**inputs)
+    
+    return speech_values
 
 def speech_to_hindi_translation(audio):
-    translated_text = translate(audio)
-    synthesised_speech = synthesise(translated_text)
+    english_text = translate_to_english(audio)
+    hindi_text = translate_english_to_hindi(english_text)
+    synthesised_speech = synthesise(hindi_text)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
-    return 16000, synthesised_speech
+    return 22050, synthesised_speech
 
 
-title = "Cascaded STST"
+title = "Speech-To-Speech-Translation for Hindi"
 description = """
-Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
-[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
-
 ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
 """