Update app.py

app.py

@@ -1,49 +1,59 @@
-import gradio as gr
-import numpy as np
 import torch
-from datasets import load_dataset
+from transformers import pipeline, VitsModel, VitsTokenizer
+import numpy as np
+import gradio as gr
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
+# Load Whisper-base
+pipe = pipeline("automatic-speech-recognition",
+                model="openai/whisper-base",
+                device=device
+)
 
-device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
-# load speech translation checkpoint
-asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
+# Define a function to translate an audio, in French here
+def translate(audio):
+    outputs = pipe(audio, max_new_tokens=256,
+                   generate_kwargs={"task": "transcribe", "language": "fr"})
+    return outputs["text"]
 
-# load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+# Load the model checkpoint and tokenizer
+model = VitsModel.from_pretrained("Matthijs/mms-tts-fra")
+tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-fra")
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
+# Define function to generate the waveform output
+def synthesise(text):
+    inputs = tokenizer(text, return_tensors="pt")
+    input_ids = inputs["input_ids"]
 
-def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
-    return outputs["text"]
+    with torch.no_grad(): 
+      outputs = model(input_ids)
+    
+    return outputs.audio[0]
 
 
-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()
+# Define global variables
+target_dtype = np.int16 # format expected by Gradio
+max_range = np.iinfo(target_dtype).max
 
 
+# Define the pipeline
 def speech_to_speech_translation(audio):
     translated_text = translate(audio)
     synthesised_speech = synthesise(translated_text)
-    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
+    synthesised_speech = (
+        synthesised_speech.numpy() * max_range).astype(target_dtype)
     return 16000, synthesised_speech
 
 
+# Define the title etc
 title = "Cascaded STST"
 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:
-
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in French. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Facebook's
+[MMS TTS](https://huggingface.co/facebook/mms-tts) model, finetuned by [Matthijs](https://huggingface.co/Matthijs), 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")
 """
 
@@ -69,4 +79,4 @@ file_translate = gr.Interface(
 with demo:
     gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
 
-demo.launch()
+demo.launch()