File size: 7,919 Bytes
44c93e2
0b6606d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
44c93e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a47d925
 
0b6606d
 
 
 
 
 
 
 
 
 
 
 
 
44c93e2
0b6606d
 
 
 
 
44c93e2
0b6606d
 
 
 
 
 
a47d925
44c93e2
 
 
 
 
 
 
 
 
 
 
 
 
 
a47d925
 
44c93e2
a47d925
 
 
44c93e2
a47d925
 
 
 
 
 
 
 
44c93e2
0b6606d
 
44c93e2
 
 
 
0b6606d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a47d925
0b6606d
a47d925
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0b6606d
 
 
 
a47d925
 
 
 
 
 
 
 
 
 
 
 
 
0b6606d
 
 
 
3f5fedd
 
 
 
 
 
 
0b6606d
 
 
 
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
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
# Environment settings
import os
os.environ["HF_HOME"] = "/tmp"
os.environ["TRANSFORMERS_CACHE"] = "/tmp"
os.environ["TORCH_HOME"] = "/tmp"
os.environ["XDG_CACHE_HOME"] = "/tmp"

import io
import re
import math
import numpy as np
import scipy.io.wavfile
import torch
from fastapi import FastAPI, Query
from fastapi.responses import StreamingResponse
from pydantic import BaseModel
from transformers import VitsModel, AutoTokenizer

app = FastAPI()

model = VitsModel.from_pretrained("Somali-tts/somali_tts_model")
tokenizer = AutoTokenizer.from_pretrained("saleolow/somali-mms-tts")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

number_words = {
    0: "eber", 1: "koow", 2: "labo", 3: "seddex", 4: "afar", 5: "shan",
    6: "lix", 7: "todobo", 8: "sideed", 9: "sagaal", 10: "toban",
    11: "toban iyo koow", 12: "toban iyo labo", 13: "toban iyo seddex",
    14: "toban iyo afar", 15: "toban iyo shan", 16: "toban iyo lix",
    17: "toban iyo todobo", 18: "toban iyo sideed", 19: "toban iyo sagaal",
    20: "labaatan", 30: "sodon", 40: "afartan", 50: "konton",
    60: "lixdan", 70: "todobaatan", 80: "sideetan", 90: "sagaashan",
    100: "boqol", 1000: "kun"
}

shortcut_map = {
    "asc": "asalaamu caleykum",
    "wcs": "wacaleykum salaam",
    "fcn": "fiican",
    "xld": "xaaladda ka waran",
    "kwrn": "kawaran",
    "scw": "salalaahu caleyhi wa salam",
    "alx": "alxamdu lilaahi",
    "m.a": "maasha allah",
    "sthy": "side tahey",
    "sxp": "saaxiib"
}

country_map = {
    "somalia": "Soomaaliya",
    "ethiopia": "Itoobiya",
    "kenya": "Kenya",
    "djibouti": "Jabuuti",
    "sudan": "Suudaan",
    "Yeman": "yemaan",
    "uganda": "Ugaandha",
    "tanzania": "Tansaaniya",
    "egypt": "Masar",
    "libya": "Liibiya",
    "algeria": "Aljeeriya",
    "morocco": "Morooko",
    "tunisia": "Tuniisiya",
    "eritrea": "Eriteriya",
    "malawi": "Malaawi",
    "English": "ingiriis",
    "Spain": "isbeen",
    "Brazil": "baraasiil",
    "niger": "Niyjer",
    "Italy": "itaaliya",
    "united states": "Maraykanka",
    "china": "Shiinaha",
    "india": "Hindiya",
    "russia": "Ruushka",
    "Saudi Arabia": "Sucuudi Carabiya",
    "germany": "Jarmalka",
    "france": "Faransiiska",
    "japan": "Jabaan",
    "canada": "Kanada",
    "australia": "Australia"
}

def number_to_words(number):
    number = int(number)
    if number < 20:
        return number_words[number]
    elif number < 100:
        tens, unit = divmod(number, 10)
        return number_words[tens * 10] + (" iyo " + number_words[unit] if unit else "")
    elif number < 1000:
        hundreds, remainder = divmod(number, 100)
        part = (number_words[hundreds] + " boqol") if hundreds > 1 else "boqol"
        if remainder:
            part += " iyo " + number_to_words(remainder)
        return part
    elif number < 1000000:
        thousands, remainder = divmod(number, 1000)
        words = [number_to_words(thousands) + " kun" if thousands > 1 else "kun"]
        if remainder:
            words.append("iyo " + number_to_words(remainder))
        return " ".join(words)
    elif number < 1000000000:
        millions, remainder = divmod(number, 1000000)
        words = [number_to_words(millions) + " milyan" if millions > 1 else "milyan"]
        if remainder:
            words.append(number_to_words(remainder))
        return " ".join(words)
    else:
        return str(number)

def normalize_text(text):
    text = re.sub(r'(?i)(?<!\w)zamzam(?!\w)', 'samsam', text)

    def replace_shortcuts(match):
        word = match.group(0).lower()
        return shortcut_map.get(word, word)
    pattern = re.compile(r'\b(' + '|'.join(re.escape(k) for k in shortcut_map.keys()) + r')\b', re.IGNORECASE)
    text = pattern.sub(replace_shortcuts, text)

    def replace_countries(match):
        word = match.group(0).lower()
        return country_map.get(word, word)
    country_pattern = re.compile(r'\b(' + '|'.join(re.escape(k) for k in country_map.keys()) + r')\b', re.IGNORECASE)
    text = country_pattern.sub(replace_countries, text)

    text = re.sub(r'(\d{1,3})(,\d{3})+', lambda m: m.group(0).replace(",", ""), text)
    text = re.sub(r'\.\d+', '', text)

    def replace_num(match):
        return number_to_words(match.group())
    text = re.sub(r'\d+', replace_num, text)

    symbol_map = {
        '$': 'doolar',
        '=': 'egwal',
        '+': 'balaas',
        '#': 'haash'
    }
    for sym, word in symbol_map.items():
        text = text.replace(sym, ' ' + word + ' ')

    text = text.replace("KH", "qa").replace("Z", "S")
    text = text.replace("SH", "SHa'a").replace("DH", "Dha'a")

    if re.search(r'(?i)(zamzam|samsam)[\s\.,!?]*$', text.strip()):
        text += " m"

    return text

def waveform_to_wav_bytes(waveform: torch.Tensor, sample_rate: int = 22050) -> bytes:
    np_waveform = waveform.cpu().numpy()
    if np_waveform.ndim == 3:
        np_waveform = np_waveform[0]
    if np_waveform.ndim == 2:
        np_waveform = np_waveform.mean(axis=0)
    np_waveform = np.clip(np_waveform, -1.0, 1.0).astype(np.float32)
    pcm_waveform = (np_waveform * 32767).astype(np.int16)
    buf = io.BytesIO()
    scipy.io.wavfile.write(buf, rate=sample_rate, data=pcm_waveform)
    buf.seek(0)
    return buf.read()

class TextIn(BaseModel):
    inputs: str

@app.post("/synthesize")
async def synthesize_post(data: TextIn):
    paragraphs = [p.strip() for p in data.inputs.split('\n') if p.strip()]
    sample_rate = getattr(model.config, "sampling_rate", 22050)
    all_waveforms = []

    for paragraph in paragraphs:
        normalized = normalize_text(paragraph)
        inputs = tokenizer(normalized, return_tensors="pt").to(device)
        with torch.no_grad():
            output = model(**inputs)
        waveform = (
            output.waveform if hasattr(output, "waveform") else
            output["waveform"] if isinstance(output, dict) and "waveform" in output else
            output[0] if isinstance(output, (tuple, list)) else
            None
        )
        if waveform is None:
            continue
        all_waveforms.append(waveform)
        silence = torch.zeros(1, sample_rate).to(waveform.device)
        all_waveforms.append(silence)

    if not all_waveforms:
        return {"error": "No audio generated."}

    final_waveform = torch.cat(all_waveforms, dim=-1)
    wav_bytes = waveform_to_wav_bytes(final_waveform, sample_rate=sample_rate)
    return StreamingResponse(io.BytesIO(wav_bytes), media_type="audio/wav")

@app.get("/synthesize")
async def synthesize_get(text: str = Query(..., description="Text to synthesize"), test: bool = Query(False)):
    if test:
        paragraphs = text.count("\n") + 1
        duration_s = paragraphs * 6
        sample_rate = 22050
        t = np.linspace(0, duration_s, int(sample_rate * duration_s), endpoint=False)
        freq = 440
        waveform = 0.5 * np.sin(2 * math.pi * freq * t).astype(np.float32)
        pcm_waveform = (waveform * 32767).astype(np.int16)
        buf = io.BytesIO()
        scipy.io.wavfile.write(buf, rate=sample_rate, data=pcm_waveform)
        buf.seek(0)
        return StreamingResponse(buf, media_type="audio/wav")

    normalized = normalize_text(text)
    inputs = tokenizer(normalized, return_tensors="pt").to(device)
    with torch.no_grad():
        output = model(**inputs)
    waveform = (
        output.waveform if hasattr(output, "waveform") else
        output["waveform"] if isinstance(output, dict) and "waveform" in output else
        output[0] if isinstance(output, (tuple, list)) else
        None
    )
    if waveform is None:
        return {"error": "Waveform not found in model output"}
    sample_rate = getattr(model.config, "sampling_rate", 22050)
    wav_bytes = waveform_to_wav_bytes(waveform, sample_rate=sample_rate)
    return StreamingResponse(io.BytesIO(wav_bytes), media_type="audio/wav")