Create handler.py
Browse files- handler.py +61 -0
handler.py
ADDED
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import time
|
2 |
+
from TTS.api import TTS
|
3 |
+
from TTS.utils.manage import ModelManager
|
4 |
+
from TTS.utils.generic_utils import get_user_data_dir
|
5 |
+
import torch
|
6 |
+
import os
|
7 |
+
from TTS.tts.configs.xtts_config import XttsConfig
|
8 |
+
import torchaudio
|
9 |
+
from TTS.tts.models.xtts import Xtts
|
10 |
+
import io
|
11 |
+
import base64
|
12 |
+
|
13 |
+
|
14 |
+
class EndpointHandler:
|
15 |
+
def __init__(self, path=""):
|
16 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
17 |
+
config = XttsConfig()
|
18 |
+
config.load_json("./model/config.json")
|
19 |
+
model = Xtts.init_from_config(config)
|
20 |
+
model.load_checkpoint(
|
21 |
+
config,
|
22 |
+
checkpoint_path="./model/model.pth",
|
23 |
+
vocab_path="./model/vocab.json",
|
24 |
+
speaker_file_path="./model/speakers_xtts.pth",
|
25 |
+
eval=True,
|
26 |
+
use_deepspeed=device == "cuda",
|
27 |
+
)
|
28 |
+
model.to(device)
|
29 |
+
|
30 |
+
self.model = model
|
31 |
+
|
32 |
+
def __call__(self, model_input):
|
33 |
+
|
34 |
+
(
|
35 |
+
gpt_cond_latent,
|
36 |
+
speaker_embedding,
|
37 |
+
) = self.model.get_conditioning_latents(
|
38 |
+
audio_path="attenborough.mp3",
|
39 |
+
gpt_cond_len=30,
|
40 |
+
gpt_cond_chunk_len=4,
|
41 |
+
max_ref_length=60,
|
42 |
+
)
|
43 |
+
|
44 |
+
print("Generating audio")
|
45 |
+
t0 = time.time()
|
46 |
+
out = self.model.inference(
|
47 |
+
text=model_input["text"],
|
48 |
+
speaker_embedding=speaker_embedding,
|
49 |
+
gpt_cond_latent=gpt_cond_latent,
|
50 |
+
temperature=0.75,
|
51 |
+
repetition_penalty=2.5,
|
52 |
+
language="en",
|
53 |
+
enable_text_splitting=True,
|
54 |
+
)
|
55 |
+
print(f"I: Time to generate audio: {inference_time} seconds")
|
56 |
+
audio_file = io.BytesIO()
|
57 |
+
torchaudio.save(audio_file, torch.tensor(out["wav"]).unsqueeze(0), 24000)
|
58 |
+
inference_time = time.time() - t0
|
59 |
+
audio_str = base64.b64encode(audio_file.getvalue()).decode("utf-8")
|
60 |
+
return {"data": audio_str, "format": "wav"}
|
61 |
+
|