import gradio as gr import numpy as np import torch from datasets import load_dataset from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline 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 tts_checkpoint = "sanchit-gandhi/speecht5_tts_vox_nl" processor = SpeechT5Processor.from_pretrained(tts_checkpoint) model = SpeechT5ForTextToSpeech.from_pretrained(tts_checkpoint).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) def translate(audio): # Trick Whisper to translate from any language to Dutch. # Note that using task=translate will translate to English instead. outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "dutch"}) return outputs["text"] def synthesise(text): # Need to specific truncate to max text positions. # Otherwise model.generate_speech will throw errors. inputs = processor( text=text, # max_length=200, max_length=598, truncation=True, # padding=True, return_tensors="pt" ) # 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() def speech_to_speech_translation(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 Dutch. Demo uses OpenAI's [Whisper Tiny](https://huggingface.co/openai/whisper-tiny) model for speech translation, and a finetuned [SpeechT5 TTS](https://huggingface.co/sanchit-gandhi/speecht5_tts_vox_nl) 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"), outputs=gr.Audio(label="Generated Speech", type="numpy"), title=title, description=description, ) file_translate = gr.Interface( fn=speech_to_speech_translation, inputs=gr.Audio(source="upload", type="filepath"), outputs=gr.Audio(label="Generated Speech", type="numpy"), examples=[["./example.wav"]], title=title, description=description, ) with demo: gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"]) demo.launch()