|
import gradio as gr |
|
import librosa |
|
import numpy as np |
|
import torch |
|
import requests |
|
|
|
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan |
|
|
|
checkpoint = "microsoft/speecht5_tts" |
|
processor = SpeechT5Processor.from_pretrained(checkpoint) |
|
model = SpeechT5ForTextToSpeech.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", |
|
"KSP": "spkemb/cmu_us_ksp_arctic-wav-arctic_b0087.npy", |
|
"RMS": "spkemb/cmu_us_rms_arctic-wav-arctic_b0353.npy", |
|
"SLT": "spkemb/cmu_us_slt_arctic-wav-arctic_a0508.npy", |
|
} |
|
|
|
def getNews(search_key): |
|
return requests.get ("https://newsapi.org/v2/everything?q=" +search_key+ "&pagesize=3&apiKey=3bca07c913ec4703a23f6ba03e15b30b").content.decode("utf-8") |
|
|
|
def getHeadlines(): |
|
return requests.get ("https://newsapi.org/v2/top-headlines?country=us&apiKey=3bca07c913ec4703a23f6ba03e15b30b").content.decode("utf-8") |
|
|
|
|
|
def predict(text, preset): |
|
if len(text.strip()) == 0: |
|
return (16000, np.zeros(0).astype(np.int16)) |
|
|
|
|
|
|
|
inputs = processor(text=getNews(text), return_tensors="pt") |
|
|
|
|
|
input_ids = inputs["input_ids"] |
|
input_ids = input_ids[..., :model.config.max_text_positions] |
|
|
|
speaker_embedding = np.load('spkemb/cmu_us_awb_arctic-wav-arctic_a0002.npy', allow_pickle=True) |
|
speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0) |
|
speech = model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder) |
|
|
|
speech = (speech.numpy() * 32767).astype(np.int16) |
|
return (16000, speech) |
|
|
|
|
|
title = "SpeechT5: Speech Synthesis" |
|
|
|
description = """ |
|
The <b>SpeechT5</b> 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. |
|
""" |
|
|
|
article = """ |
|
<div style='margin:20px auto;'> |
|
|
|
<p>References: <a href="https://arxiv.org/abs/2110.07205">SpeechT5 paper</a> | |
|
<a href="https://github.com/microsoft/SpeechT5/">original GitHub</a> | |
|
<a href="https://huggingface.co/mechanicalsea/speecht5-tts">original weights</a></p> |
|
|
|
<p>Speaker embeddings were generated from <a href="http://www.festvox.org/cmu_arctic/">CMU ARCTIC</a> using <a href="https://huggingface.co/mechanicalsea/speecht5-vc/blob/main/manifest/utils/prep_cmu_arctic_spkemb.py">this script</a>.</p> |
|
|
|
</div> |
|
""" |
|
|
|
examples = [ |
|
["example 1", "US"], |
|
["example 2", "US"], |
|
] |
|
|
|
gr.Interface( |
|
fn=predict, |
|
inputs=[ |
|
gr.Text(label="Input Text"), |
|
gr.Radio(label="Preset", choices=[ |
|
"US", |
|
"International", |
|
"Technology", |
|
"KPop", |
|
"Surprise Me!" |
|
], value="KPop"), |
|
], |
|
outputs=[ |
|
gr.Audio(label="Generated Speech", type="numpy"), |
|
], |
|
title=title, |
|
description=description, |
|
article=article, |
|
examples=examples, |
|
).launch(share=True) |
|
|
