import gradio as gr import librosa import numpy as np import torch from transformers import SpeechT5Processor, SpeechT5ForSpeechToSpeech, SpeechT5HifiGan checkpoint = "microsoft/speecht5_vc" processor = SpeechT5Processor.from_pretrained(checkpoint) model = SpeechT5ForSpeechToSpeech.from_pretrained(checkpoint) vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan") speaker_embeddings = { "BDL": "spkemb/cmu_us_bdl_arctic-wav-arctic_a0009.npy", "CLB": "spkemb/cmu_us_clb_arctic-wav-arctic_a0144.npy", "RMS": "spkemb/cmu_us_rms_arctic-wav-arctic_b0353.npy", "SLT": "spkemb/cmu_us_slt_arctic-wav-arctic_a0508.npy", } def process_audio(sampling_rate, waveform): # convert from int16 to floating point waveform = waveform / 32678.0 # convert to mono if stereo if len(waveform.shape) > 1: waveform = librosa.to_mono(waveform.T) # resample to 16 kHz if necessary if sampling_rate != 16000: waveform = librosa.resample(waveform, orig_sr=sampling_rate, target_sr=16000) # limit to 30 seconds waveform = waveform[:16000*30] # make PyTorch tensor waveform = torch.tensor(waveform) return waveform def predict(speaker, audio, mic_audio=None): # audio = tuple (sample_rate, frames) or (sample_rate, (frames, channels)) if mic_audio is not None: sampling_rate, waveform = mic_audio elif audio is not None: sampling_rate, waveform = audio else: return (16000, np.zeros(0).astype(np.int16)) waveform = process_audio(sampling_rate, waveform) inputs = processor(audio=waveform, sampling_rate=16000, return_tensors="pt") speaker_embedding = np.load(speaker_embeddings[speaker[:3]]) speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0) speech = model.generate_speech(inputs["input_values"], speaker_embedding, vocoder=vocoder) speech = (speech.numpy() * 32767).astype(np.int16) return (16000, speech) title = "SpeechT5: Voice Conversion" description = """ The SpeechT5 model is pre-trained on text as well as speech inputs, with targets that are also a mix of text and speech. By pre-training on text and speech at the same time, it learns unified representations for both, resulting in improved modeling capabilities. SpeechT5 can be fine-tuned for different speech tasks. This space demonstrates the speech-to-speech checkpoint for (American) English language voice conversion. See also the speech recognition (ASR) demo and the text-to-speech (TTS) demo. How to use: Upload an audio file or record using the microphone. The audio is converted to mono and resampled to 16 kHz before being passed into the model. The output is a mel spectrogram, which is converted to a mono 16 kHz waveform by the HiFi-GAN vocoder. Because the model always applies random dropout, each attempt will give slightly different results. """ article = """

References: SpeechT5 paper | original GitHub | original weights

@article{Ao2021SpeechT5,
  title   = {SpeechT5: Unified-Modal Encoder-Decoder Pre-training for Spoken Language Processing},
  author  = {Junyi Ao and Rui Wang and Long Zhou and Chengyi Wang and Shuo Ren and Yu Wu and Shujie Liu and Tom Ko and Qing Li and Yu Zhang and Zhihua Wei and Yao Qian and Jinyu Li and Furu Wei},
  eprint={2110.07205},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  year={2021}
}

Example sound credits:

Speaker embeddings were generated from CMU ARCTIC using this script.

""" examples = [ ["BDL (male)", "examples/yearn_for_time.mp3", None], ["CLB (female)", "examples/henry5.mp3", None], ["RMS (male)", "examples/see_in_eyes.wav", None], ["SLT (female)", "examples/hmm_i_dont_know.wav", None], ] gr.Interface( fn=predict, inputs=[ gr.Radio(label="Speaker", choices=["BDL (male)", "CLB (female)", "RMS (male)", "SLT (female)"], value="BDL (male)"), gr.Audio(label="Upload Speech", source="upload", type="numpy"), gr.Audio(label="Record Speech", source="microphone", type="numpy"), ], outputs=[ gr.Audio(label="Converted Speech", type="numpy"), ], title=title, description=description, article=article, examples=examples, #cache_examples=True, ).launch()