Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import re
|
3 |
+
import uuid
|
4 |
+
import torch
|
5 |
+
import torchaudio
|
6 |
+
import soundfile as sf
|
7 |
+
from fastapi import FastAPI
|
8 |
+
from fastapi.responses import FileResponse
|
9 |
+
from pydantic import BaseModel
|
10 |
+
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
|
11 |
+
from speechbrain.inference.speaker import EncoderClassifier
|
12 |
+
|
13 |
+
app = FastAPI()
|
14 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
15 |
+
CACHE_DIR = "/tmp/hf-cache"
|
16 |
+
|
17 |
+
# Load models
|
18 |
+
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts", cache_dir=CACHE_DIR)
|
19 |
+
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan", cache_dir=CACHE_DIR).to(device)
|
20 |
+
model_male = SpeechT5ForTextToSpeech.from_pretrained("Somalitts/5aad", cache_dir=CACHE_DIR).to(device)
|
21 |
+
model_female = SpeechT5ForTextToSpeech.from_pretrained("Somalitts/8aad", cache_dir=CACHE_DIR).to(device)
|
22 |
+
|
23 |
+
# Speaker encoder
|
24 |
+
speaker_model = EncoderClassifier.from_hparams(
|
25 |
+
source="speechbrain/spkrec-xvect-voxceleb",
|
26 |
+
run_opts={"device": device},
|
27 |
+
savedir="/tmp/spk_model"
|
28 |
+
)
|
29 |
+
|
30 |
+
# Load speaker embeddings
|
31 |
+
def get_embedding(wav_path, pt_path):
|
32 |
+
if os.path.exists(pt_path):
|
33 |
+
return torch.load(pt_path).to(device)
|
34 |
+
audio, sr = torchaudio.load(wav_path)
|
35 |
+
audio = torchaudio.functional.resample(audio, sr, 16000).mean(dim=0).unsqueeze(0).to(device)
|
36 |
+
with torch.no_grad():
|
37 |
+
emb = speaker_model.encode_batch(audio)
|
38 |
+
emb = torch.nn.functional.normalize(emb, dim=2).squeeze()
|
39 |
+
torch.save(emb.cpu(), pt_path)
|
40 |
+
return emb
|
41 |
+
|
42 |
+
embedding_male = get_embedding("Hussein.wav", "/tmp/male_embedding.pt")
|
43 |
+
embedding_female = get_embedding("caasho.wav", "/tmp/female_embedding.pt")
|
44 |
+
|
45 |
+
# Text normalization
|
46 |
+
number_words = {
|
47 |
+
0: "eber", 1: "koow", 2: "labo", 3: "seddex", 4: "afar", 5: "shan",
|
48 |
+
6: "lix", 7: "todobo", 8: "sideed", 9: "sagaal", 10: "toban",
|
49 |
+
20: "labaatan", 30: "sodon", 40: "afartan", 50: "konton",
|
50 |
+
60: "lixdan", 70: "todobaatan", 80: "sideetan", 90: "sagaashan",
|
51 |
+
100: "boqol", 1000: "kun"
|
52 |
+
}
|
53 |
+
|
54 |
+
def number_to_words(n):
|
55 |
+
if n < 20:
|
56 |
+
return number_words.get(n, str(n))
|
57 |
+
elif n < 100:
|
58 |
+
tens, unit = divmod(n, 10)
|
59 |
+
return number_words[tens * 10] + (" " + number_words[unit] if unit else "")
|
60 |
+
elif n < 1000:
|
61 |
+
hundreds, rem = divmod(n, 100)
|
62 |
+
return (number_words[hundreds] + " boqol" if hundreds > 1 else "boqol") + (" " + number_to_words(rem) if rem else "")
|
63 |
+
elif n < 1_000_000:
|
64 |
+
th, rem = divmod(n, 1000)
|
65 |
+
return (number_to_words(th) + " kun") + (" " + number_to_words(rem) if rem else "")
|
66 |
+
else:
|
67 |
+
return str(n)
|
68 |
+
|
69 |
+
def replace_numbers_with_words(text):
|
70 |
+
return re.sub(r'\b\d+\b', lambda m: number_to_words(int(m.group())), text)
|
71 |
+
|
72 |
+
def normalize_text(text):
|
73 |
+
text = text.lower()
|
74 |
+
text = replace_numbers_with_words(text)
|
75 |
+
text = re.sub(r'[^\w\s]', '', text)
|
76 |
+
return text
|
77 |
+
|
78 |
+
# API request schema
|
79 |
+
class TTSRequest(BaseModel):
|
80 |
+
text: str
|
81 |
+
voice: str # "Male" or "Female"
|
82 |
+
|
83 |
+
@app.post("/speak")
|
84 |
+
def speak(payload: TTSRequest):
|
85 |
+
clean_text = normalize_text(payload.text)
|
86 |
+
inputs = processor(text=clean_text, return_tensors="pt").to(device)
|
87 |
+
model = model_male if payload.voice.lower() == "male" else model_female
|
88 |
+
embedding = embedding_male if payload.voice.lower() == "male" else embedding_female
|
89 |
+
|
90 |
+
with torch.no_grad():
|
91 |
+
waveform = model.generate_speech(inputs["input_ids"], embedding.unsqueeze(0), vocoder=vocoder)
|
92 |
+
|
93 |
+
out_path = f"/tmp/{uuid.uuid4().hex}.wav"
|
94 |
+
sf.write(out_path, waveform.cpu().numpy(), 16000)
|
95 |
+
return FileResponse(out_path, media_type="audio/wav", filename="voice.wav")
|