Update for Russian app.py

app.py

@@ -1,49 +1,76 @@
+"Задание 3"
+
+!pip install datasets  #>=2.6.1
+!pip install git+https://github.com/huggingface/transformers
+!pip -q install gradio==3.28.3
+!pip install phonemizer
+!pip install py-espeak-ng
+
+%pip install sentencepiece
+%pip install --pre ekorpkit
+
 import gradio as gr
 import numpy as np
 import torch
+import espeakng
+import sentencepiece
 from datasets import load_dataset
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
-
+from transformers import pipeline, MarianMTModel, MarianTokenizer, VitsModel, VitsTokenizer
 
 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)
-
-# load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+import phonemizer
+model_wav2vec = 'facebook/wav2vec2-lv-60-espeak-cv-ft'
+asr_pipe = pipeline("automatic-speech-recognition", model=model_wav2vec, device=device)
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").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)
+# from speech to text
+def translate_audio(audio):
+    outputs = asr_pipe(audio, max_new_tokens=256, 
+                       generate_kwargs={"task": "translate"})
+    return outputs["text"]
 
+# translation
+def translate_text(text, from_language, target_language):  #to English -mul en, to Russian - en ru
+    model_name = f'Helsinki-NLP/opus-mt-{from_language}-{target_language}'
+    tokenizer = MarianTokenizer.from_pretrained(model_name)
+    model = MarianMTModel.from_pretrained(model_name)
 
-def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
-    return outputs["text"]
+    inputs = tokenizer.encode(text, return_tensors="pt")
+    outputs = model.generate(inputs, num_beams=4, max_length=50, early_stopping=True)
+    translated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return translated_text
 
+# load text-to-speech checkpoint
+#model = pipeline("text-to-speech", model="voxxer/speecht5_finetuned_commonvoice_ru_translit")
+model = VitsModel.from_pretrained("voxxer/speecht5_finetuned_commonvoice_ru_translit")
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-rus")
 
 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)
+    translated_text = translate_text(text, 'mul', 'en')
+    translated_text = translate_text(translate_text, 'en', 'ru')
+    inputs = tokenizer(translated_text, return_tensors="pt")
+    input_ids = inputs["input_ids"]
+    with torch.no_grad():
+        outputs = model(input_ids)
+    speech = outputs["waveform"]
     return speech.cpu()
 
 
 def speech_to_speech_translation(audio):
-    translated_text = translate(audio)
-    synthesised_speech = synthesise(translated_text)
+    text_from_audio = translate_audio(audio)
+    print(translated_text)
+    synthesised_speech = synthesise(text_from_audio)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
-    return 16000, synthesised_speech
+    return 16000, synthesised_speech[0]
 
 
 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:
-
+* В начале происходит распознавание речи с помощью модели facebook/wav2vec2-lv-60-espeak-cv-ft и на выходе получается текст на любом из 60 языков.
+* Затем полученный текст переводится сначала на английский с помощью Helsinki-NLP/opus-mt-mul-en, а потом на русский с помощью Helsinki-NLP/opus-mt-en-ru
+* На последнем шаге полученный текст озвучивается с помощью fine-tune-говой версии microsoft/speecht5_tts - voxxer/speecht5_finetuned_commonvoice_ru_translit
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Russian. Demo uses facebook/mms-tts-rus 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")
 """