File size: 3,417 Bytes
79f8132 |
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 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 |
from typing import Dict, Any,Union
import tempfile
import numpy as np
import torch
import pyewts
import noisereduce as nr
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
from num2tib.core import convert
from num2tib.core import convert2text
import soundfile as sf
import base64
import re
import requests
import os
from pydub import AudioSegment
converter = pyewts.pyewts()
def download_file(url, destination):
response = requests.get(url)
with open(destination, 'wb') as file:
file.write(response.content)
# Example usage:
download_file('https://huggingface.co/openpecha/speecht5-tts-01/resolve/main/female_2.npy', 'female_2.npy')
def replace_numbers_with_convert(sentence, wylie=True):
pattern = r'\d+(\.\d+)?'
def replace(match):
return convert(match.group(), wylie)
result = re.sub(pattern, replace, sentence)
return result
def cleanup_text(inputs):
for src, dst in replacements:
inputs = inputs.replace(src, dst)
return inputs
speaker_embeddings = {
"Lhasa(female)": "female_2.npy",
}
replacements = [
('_', '_'),
('*', 'v'),
('`', ';'),
('~', ','),
('+', ','),
('\\', ';'),
('|', ';'),
('╚',''),
('╗','')
]
class EndpointHandler():
def __init__(self, path=""):
# load the model
self.processor = SpeechT5Processor.from_pretrained("TenzinGayche/TTS_run3_ep20_174k_b")
self.model = SpeechT5ForTextToSpeech.from_pretrained("TenzinGayche/TTS_run3_ep20_174k_b")
self.model.to('cuda')
self.vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
def __call__(self, data: Dict[str, Any]) -> Dict[str, Union[int, str]]:
"""_summary_
Args:
data (Dict[str, Any]): _description_
Returns:
bytes: _description_
"""
text = data.pop("inputs",data)
# process input
if len(text.strip()) == 0:
return (16000, np.zeros(0).astype(np.int16))
text = converter.toWylie(text)
text=cleanup_text(text)
text=replace_numbers_with_convert(text)
inputs = self.processor(text=text, return_tensors="pt")
input_ids = inputs["input_ids"]
input_ids = input_ids[..., :self.model.config.max_text_positions]
speaker_embedding = np.load(speaker_embeddings['Lhasa(female)'])
speaker_embedding = torch.tensor(speaker_embedding)
speech = self.model.generate_speech(input_ids.to('cuda'), speaker_embedding.to('cuda'), vocoder=self.vocoder.to('cuda'))
speech = nr.reduce_noise(y=speech.to('cpu'), sr=16000)
# Create a unique temporary WAV file
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_wav_file:
temp_wav_path = temp_wav_file.name
sf.write(temp_wav_path, speech, 16000, 'PCM_24') # Use sf.write to write the WAV file
# Read the WAV file and encode it as base64
with open(temp_wav_path, "rb") as wav_file:
audio_base64 = base64.b64encode(wav_file.read()).decode("utf-8")
# Clean up the temporary WAV file
os.remove(temp_wav_path)
return {
"sample_rate": 16000,
"audio_base64": audio_base64, # Base64-encoded audio data
"model": "openpecha/speecht5-tts-01",
"model_version": "1"
}
|