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 # 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) processor = SpeechT5Processor.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl") model = SpeechT5ForTextToSpeech.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl").to(device) vocoder = SpeechT5HifiGan.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl").to(device) # embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation") # speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0) dataset_nl = load_dataset("facebook/voxpopuli", "nl", split="train", streaming=True) data_list = [] speaker_embeddings_list = [] for i, data in enumerate(iter(dataset_nl)): # print(i) if(i > 5): break data_list.append(data) # data = next(iter(dataset_nl)) text = data["raw_text"] # print(data) speaker_embeddings = create_speaker_embedding(data["audio"]["array"]) speaker_embeddings = torch.tensor(speaker_embeddings)[None] speaker_embeddings_list.append(speaker_embeddings) speaker_embeddings = speaker_embeddings_list[4] def translate(audio): # outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"}) outputs = pipe(audio, max_new_tokens=256, generate_kwargs={"language":"<|nl|>","task": "transcribe"}) return outputs["text"] 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() 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 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: ![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()