File size: 2,109 Bytes
32e218a
b84dff6
32e218a
b84dff6
32e218a
b84dff6
 
 
 
2176d40
 
b84dff6
 
 
2176d40
 
 
b84dff6
e30c131
2176d40
b84dff6
 
 
2176d40
b84dff6
32e218a
2176d40
b84dff6
 
 
f45d8c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
import gradio as gr
# from TTS.api import TTS

# tts = TTS(model_name="tts_models/multilingual/multi-dataset/your_tts", progress_bar=False, gpu=False)

# def predict(text):
#     file_path = "output.wav"
#     tts.tts_to_file(text, speaker=tts.speakers[0], language="en", file_path=file_path)
#     return file_path

# demo = gr.Interface(
#     fn=predict,
#     inputs='text',
#     outputs='audio'
# )


# demo.launch()


import librosa
import numpy as np
import torch

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")

def predict(text):
    if len(text.strip()) == 0:
        return (16000, np.zeros(0).astype(np.int16))

    inputs = processor(text=text, return_tensors="pt")

    # limit input length
    input_ids = inputs["input_ids"]
    input_ids = input_ids[..., :model.config.max_text_positions]

    # if speaker == "Surprise Me!":
    #     # load one of the provided speaker embeddings at random
    #     idx = np.random.randint(len(speaker_embeddings))
    #     key = list(speaker_embeddings.keys())[idx]
    #     speaker_embedding = np.load(speaker_embeddings[key])

    #     # randomly shuffle the elements
    #     np.random.shuffle(speaker_embedding)

    #     # randomly flip half the values
    #     x = (np.random.rand(512) >= 0.5) * 1.0
    #     x[x == 0] = -1.0
    #     speaker_embedding *= x

        #speaker_embedding = np.random.rand(512).astype(np.float32) * 0.3 - 0.15
    # else:
    speaker_embedding = np.load("cmu_us_bdl_arctic-wav-arctic_a0009.npy")

    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)

demo = gr.Interface(
    fn = predict,
    inputs="text",
    outputs=gr.Audio(type="numpy")
)

demo.launch()