Update app.py

app.py

@@ -1,82 +1,63 @@
 import gradio as gr
 import numpy as np
 import torch
-from transformers import AutoProcessor, pipeline, BarkModel, GenerationConfig
+from transformers import AutoProcessor, pipeline, VitsModel, VitsTokenizer
 
-ASR_MODEL_NAME = "bofenghuang/whisper-large-v2-cv11-german"
-TTS_MODEL_NAME = "suno/bark-small"
-BATCH_SIZE = 8
-voices = {
-"male" : "v2/de_speaker_0",
-"female" : "v2/de_speaker_3"
-}
-
-device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 # load speech translation checkpoint
+ASR_MODEL_NAME = 'openai/whisper-base'
 asr_pipe = pipeline("automatic-speech-recognition", model=ASR_MODEL_NAME, chunk_length_s=30,device=device)
 
-MULTILINGUAL = True  # set True for multilingual models, False for English-only
-
-if MULTILINGUAL:
-    generation_config = GenerationConfig.from_pretrained("openai/whisper-large-v2")
 
 # load text-to-speech checkpoint
-processor = AutoProcessor.from_pretrained("suno/bark-small")
-model = BarkModel.from_pretrained("suno/bark-small").to(device)
-sampling_rate = model.generation_config.sample_rate
+model = VitsModel.from_pretrained("Matthijs/mms-tts-deu")
+tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-deu")
 
-# set the forced ids
-model.config.forced_decoder_ids = None
-model.config.suppress_tokens = []
 
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "de"})
     return outputs["text"]
     
 def synthesise(text, voice_preset):
-    inputs = processor(text=text, return_tensors="pt",voice_preset=voice_preset)
-    speech = model.generate(**inputs.to(device))
-    return speech[0]
+    inputs = tokenizer(text_example, return_tensors="pt")
+    input_ids = inputs["input_ids"]
+
+    with torch.no_grad():
+        outputs = model(input_ids)
     
-def speech_to_speech_translation(audio, voice):
-    voice_preset = None
+    speech = outputs.audio[0]
+    return speech.cpu()
+    
+def speech_to_speech_translation(audio):
     translated_text = translate(audio)
-    print(translated_text)
-    if voice == "Female":
-     voice_preset = voices["female"]
-    else:
-     voice_preset = voices["male"]
-    synthesised_speech = synthesise(translated_text, voice_preset)
-    synthesised_speech = (synthesised_speech.cpu().numpy() * 32767).astype(np.int16)
-    return sampling_rate, synthesised_speech
+    synthesised_speech = synthesise(translated_text)
+    synthesised_speech = (synthesised_speech.numpy() * max_range).astype(np.int16)
+    return 16000, synthesised_speech
+    
     
 title = "Cascaded STST - Any language to German speech"
 description = """
-Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in German. Demo uses fine-tuned version of openai/whisper-large-v2 model (https://huggingface.co/bofenghuang/whisper-large-v2-cv11-german) for speech translation, and Suno's
-[Bark-large](https://huggingface.co/suno/bark-small) model for text-to-speech:
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in German. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
+[MMS TTS](https://huggingface.co/Matthijs/mms-tts-deu) 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")
 """
 demo = gr.Blocks()
 
 mic_translate = gr.Interface(
     fn=speech_to_speech_translation,
-    inputs=[gr.Audio(source="microphone", type="filepath"),
-    gr.inputs.Radio(["Male", "Female"], label="Voice", default="Male")],
+    inputs=gr.Audio(source="microphone", type="filepath"),
     outputs=gr.Audio(label="Generated Speech", type="numpy"),
     title=title,
     description=description,
-    allow_flagging="never"
 )
 
 file_translate = gr.Interface(
     fn=speech_to_speech_translation,
-    inputs=[gr.Audio(source="upload", type="filepath"),
-    gr.inputs.Radio(["Male", "Female"], label="Voice", default="Male")],
+    inputs=gr.Audio(source="upload", type="filepath"),
     outputs=gr.Audio(label="Generated Speech", type="numpy"),
+    examples=[["./example.wav"]],
     title=title,
     description=description,
-    allow_flagging="never"
 )
 
 with demo: