Update app.py

app.py

@@ -1,19 +1,16 @@
-import subprocess
-
-# Установка зависимостей внутри пространства
-subprocess.run(["pip", "install", "vosk"])
-subprocess.run(["pip", "install", "SpeechRecognition"])
-
 import gradio as gr
 import numpy as np
 import torch
-import speech_recognition as sr
 from datasets import load_dataset
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor
+
+from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
-# Load a lightweight text-to-speech checkpoint and speaker embeddings
+# load speech translation checkpoint
+asr_pipe = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-base-960h", device=device)
+
+# load text-to-speech checkpoint and speaker embeddings
 processor = SpeechT5Processor.from_pretrained("ttskit/ttskit-tts-ljspeech")
 
 model = SpeechT5ForTextToSpeech.from_pretrained("ttskit/ttskit-tts-ljspeech").to(device)
@@ -22,11 +19,9 @@ vocoder = SpeechT5HifiGan.from_pretrained("ljspeech/vocoder-cryptron").to(device
 embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
 speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
-def recognize_speech(audio):
-    recognizer = sr.Recognizer()
-    with sr.AudioFile(audio) as source:
-        audio_data = recognizer.record(source)
-    return recognizer.recognize_google(audio_data)
+def translate(audio):
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
+    return outputs["text"]
 
 def synthesise(text):
     inputs = processor(text=text, return_tensors="pt")
@@ -34,14 +29,14 @@ def synthesise(text):
     return speech.cpu()
 
 def speech_to_speech_translation(audio):
-    translated_text = recognize_speech(audio)
+    translated_text = translate(audio)
     synthesised_speech = synthesise(translated_text)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
     return 16000, synthesised_speech
 
 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 Google Web Speech API for automatic speech recognition, and lightweight text-to-speech and vocoder models.
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses Facebook's [Wav2Vec 2.0](https://huggingface.co/facebook/wav2vec2-base-960h) model for speech recognition, and a lightweight text-to-speech model ([ttskit/ttskit-tts-ljspeech](https://huggingface.co/ttskit/ttskit-tts-ljspeech)) along with a lightweight vocoder ([ljspeech/vocoder-cryptron](https://huggingface.co/ljspeech/vocoder-cryptron)).
 """
 
 demo = gr.Blocks()
@@ -66,4 +61,4 @@ file_translate = gr.Interface(
 with demo:
     gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
 
-demo.launch()
+demo.launch()