ememzyvisuals's picture
Upload app.py
001420f verified
Raw History Blame Contribute Delete
11 kB
import os
import torch
import torchaudio as ta
import tempfile
import gradio as gr
import spaces
from huggingface_hub import hf_hub_download
MODEL_REPO = "Axiveri/WazobiaVoice"
from wazobiavoice_tts.mtl_tts import WazobiaVoiceMultilingualTTS
print("Loading WazobiaVoice on CPU (moved to GPU per-request via ZeroGPU) ...")
model = WazobiaVoiceMultilingualTTS.from_pretrained("cpu")
print("Model ready.")
# ── Personas ─────────────────────────────────────────────────────────────────
# display name -> (language_id, inbuilt_voices reference filename)
VOICES = {
"Wura (Yoruba, F)": ("yo", "wura.wav"),
"Bọlaji (Yoruba, M)": ("yo", "bolaji.wav"),
"Elder Deji (Yoruba, M)": ("yo", "elder_deji.wav"),
"IfΓ© (Yoruba, F)": ("yo", "ife.wav"),
"Musa (Hausa, M)": ("ha", "musa.wav"),
"Hauwa (Hausa, F)": ("ha", "hauwa.wav"),
"Emeka (Igbo, M)": ("ig", "emeka.wav"),
"Adaeze (Igbo, F)": ("ig", "adaeze.wav"),
"Tunde (Pidgin, M)": ("pcm", "tunde.wav"),
"Ngozi (Pidgin, F)": ("pcm", "ngozi.wav"),
"John (Pidgin, M, teen)": ("pcm", "john.wav"),
"James (English, M)": ("en", "james.wav"),
"Amara (English, F)": ("en", "amara.wav"),
}
LANG_LABELS = {
"yo": "Yoruba", "ha": "Hausa", "ig": "Igbo", "pcm": "Nigerian Pidgin", "en": "Nigerian English",
}
REF_CACHE = {}
def get_ref_audio(fname):
if fname not in REF_CACHE:
REF_CACHE[fname] = hf_hub_download(
repo_id=MODEL_REPO, filename=f"inbuilt_voices/{fname}"
)
return REF_CACHE[fname]
EXAMPLES = [
["Ọrẹ́ mi ti dΓ© lΓ‘ti Γ¬lΓΊ αΊΉΜ€kọ́ nΓ­ Γ²nΓ­, a sΓ¬ Ε„ sọ̀rọ̀ nΓ­pa bΓ­ a αΉ£e Ε„ lọ sΓ­ ọjΓ  lΓ‘ti ra αΊΉΜ€wΓ  Γ ti αΊΉja tΓ³ dΓ‘ra fΓΊn oΓΊnjαΊΉΜ€ alẹ́.", "Wura (Yoruba, F)"],
["Yau na tashi da safe na ci abincin safe sannan na fita zuwa kasuwa don sayen abubuwa na gida kamar gero da yaji da kuma nama mai kyau.", "Musa (Hausa, M)"],
["N'α»₯tα»₯tα»₯, m gara ahα»‹a zα»₯ta ihe oriri dα»‹ ka akwα»₯kwọ nri na azα»₯Μ€ ọhα»₯rα»₯ maka nri abalα»‹, α»₯mα»₯aka na-egwu egwu n'azα»₯ α»₯lọ.", "Emeka (Igbo, M)"],
["My guy don land from Abuja since morning, we dey gist about how we go reach market buy beans and fresh fish for evening chop.", "Tunde (Pidgin, M)"],
["My friend arrived from Lagos this morning, and we've been chatting about going to the market to buy beans and fresh fish for dinner.", "James (English, M)"],
]
@spaces.GPU
def synthesize(text, voice_name, exaggeration, cfg_weight):
if not text or not text.strip():
return None, "Please enter some text."
lang_id, ref_fname = VOICES.get(voice_name, ("en", "james.wav"))
try:
model.to("cuda")
ref_path = get_ref_audio(ref_fname)
wav = model.generate(
text,
language_id=lang_id,
audio_prompt_path=ref_path,
exaggeration=exaggeration,
cfg_weight=cfg_weight,
)
tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
ta.save(tmp.name, wav, model.sr)
return tmp.name, f"Generated in {LANG_LABELS[lang_id]} using {voice_name}."
except Exception as e:
return None, f"Error: {str(e)}"
@spaces.GPU
def clone_voice(text, language_name, ref_audio, exaggeration, cfg_weight):
if not text or not text.strip():
return None, "Please enter some text."
if ref_audio is None:
return None, "Please upload or record a reference clip (5-10 seconds)."
lang_id = {v: k for k, v in LANG_LABELS.items()}[language_name]
try:
model.to("cuda")
wav = model.generate(
text,
language_id=lang_id,
audio_prompt_path=ref_audio,
exaggeration=exaggeration,
cfg_weight=cfg_weight,
)
tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
ta.save(tmp.name, wav, model.sr)
return tmp.name, f"Generated {LANG_LABELS[lang_id]} speech in the cloned voice."
except Exception as e:
return None, f"Error: {str(e)}"
# ── Theme: orange + green + white ───────────────────────────────────────────
ORANGE = gr.themes.Color(
c50="#fff7ed", c100="#ffedd5", c200="#fed7aa",
c300="#fdba74", c400="#fb923c", c500="#f97316",
c600="#ea580c", c700="#c2410c", c800="#9a3412",
c900="#7c2d12", c950="#431407",
)
GREEN = gr.themes.Color(
c50="#f0fdf4", c100="#dcfce7", c200="#bbf7d0",
c300="#86efac", c400="#4ade80", c500="#16a34a",
c600="#15803d", c700="#166534", c800="#14532d",
c900="#0f3d22", c950="#052e16",
)
theme = gr.themes.Soft(
primary_hue=ORANGE,
secondary_hue=GREEN,
).set(
button_primary_background_fill="linear-gradient(135deg,#f97316 0%,#ea580c 100%)",
button_primary_background_fill_hover="linear-gradient(135deg,#fb923c 0%,#f97316 100%)",
button_primary_text_color="#ffffff",
slider_color="#f97316",
)
CSS = """
html, body { overflow-x: hidden !important; }
.gradio-container { max-width: 900px !important; margin: 0 auto !important; padding: 0 1rem !important; overflow-x: hidden !important; box-sizing: border-box !important; }
.gradio-container * { box-sizing: border-box !important; }
/* The audio waveform (wavesurfer.js canvas) can render at a fixed pixel
width wider than a mobile viewport instead of scaling to its container,
which pushes the whole page wider than the screen -- only becoming
playable once landscape rotation makes the viewport exceed that fixed
width. Force it to always respect 100% of its container instead. */
.gradio-container [data-testid="waveform"],
.gradio-container [data-testid="waveform"] *,
.gradio-container audio {
max-width: 100% !important;
width: 100% !important;
}
.gradio-container [data-testid="waveform"] canvas {
max-width: 100% !important;
}
.wv-header { text-align: center; padding: 2rem 0 0.5rem; }
.wv-title {
font-size: clamp(2.2rem, 6vw, 3.4rem);
font-weight: 700;
letter-spacing: -0.03em;
background: linear-gradient(135deg, #16a34a 0%, #f97316 100%);
-webkit-background-clip: text;
-webkit-text-fill-color: transparent;
background-clip: text;
line-height: 1.1;
margin: 0 0 0.5rem;
}
.wv-sub { color: #15803d; font-size: 0.95rem; margin: 0 0 1.5rem; }
.wv-footer {
text-align: center;
padding: 1.5rem 0 2rem;
color: #15803d;
font-size: 0.82rem;
border-top: 1px solid #bbf7d0;
margin-top: 1.5rem;
}
.wv-footer a { color: #c2410c; text-decoration: none; }
.wv-footer a:hover { text-decoration: underline; }
"""
# ── UI ───────────────────────────────────────────────────────────────────────
with gr.Blocks(title="WazobiaVoice Demo", theme=theme, css=CSS) as demo:
gr.HTML("""
<div class="wv-header">
<h1 class="wv-title">WazobiaVoice</h1>
<p class="wv-sub">
Nigerian multilingual TTS &amp; voice cloning &nbsp;Β·&nbsp;
Yoruba &nbsp;Β·&nbsp; Hausa &nbsp;Β·&nbsp; Igbo &nbsp;Β·&nbsp; Pidgin &nbsp;Β·&nbsp; Nigerian English
</p>
</div>
""")
with gr.Tab("Synthesize"):
with gr.Row():
with gr.Column(scale=3):
text_input = gr.Textbox(
label="Text",
placeholder="Type text in the language matching your chosen voice...",
lines=3,
)
voice_dropdown = gr.Dropdown(
choices=list(VOICES.keys()),
value="IfΓ© (Yoruba, F)",
label="Voice",
)
with gr.Row():
exaggeration_slider = gr.Slider(
minimum=0.3, maximum=1.0, value=0.5, step=0.05, label="Exaggeration",
)
cfg_slider = gr.Slider(
minimum=0.2, maximum=1.0, value=0.5, step=0.05, label="CFG Weight",
)
generate_btn = gr.Button("Generate Speech", variant="primary", elem_id="generate-btn")
with gr.Column(scale=2):
audio_output = gr.Audio(label="Generated Speech", type="filepath")
status_text = gr.Textbox(label="Status", interactive=False)
gr.Examples(
examples=EXAMPLES,
inputs=[text_input, voice_dropdown],
label="Try these",
)
generate_btn.click(
fn=synthesize,
inputs=[text_input, voice_dropdown, exaggeration_slider, cfg_slider],
outputs=[audio_output, status_text],
)
with gr.Tab("Voice Cloning"):
gr.Markdown(
"Upload or record a **5-10 second** reference clip. "
"WazobiaVoice will speak your chosen language in that voice."
)
with gr.Row():
with gr.Column(scale=3):
clone_text = gr.Textbox(
label="Text to Generate",
lines=3,
)
clone_language = gr.Dropdown(
choices=list(LANG_LABELS.values()),
value="Yoruba",
label="Language",
)
clone_ref_audio = gr.Audio(
sources=["microphone", "upload"],
label="Reference Audio (5-10 seconds)",
type="filepath",
)
with gr.Row():
clone_exaggeration = gr.Slider(
minimum=0.3, maximum=1.0, value=0.5, step=0.05, label="Exaggeration",
)
clone_cfg = gr.Slider(
minimum=0.2, maximum=1.0, value=0.5, step=0.05, label="CFG Weight",
)
clone_btn = gr.Button("Clone & Generate", variant="primary", elem_id="clone-btn")
with gr.Column(scale=2):
clone_audio_output = gr.Audio(label="Cloned Speech", type="filepath")
clone_status = gr.Textbox(label="Status", interactive=False)
clone_btn.click(
fn=clone_voice,
inputs=[clone_text, clone_language, clone_ref_audio, clone_exaggeration, clone_cfg],
outputs=[clone_audio_output, clone_status],
)
gr.HTML("""
<div class="wv-footer">
<a href="https://huggingface.co/Axiveri/WazobiaVoice">Model card</a>
&nbsp;Β·&nbsp;
<a href="https://github.com/Ememzyvisuals/wazobiavoice-TTS">GitHub</a>
&nbsp;Β·&nbsp;
Built by <a href="https://ememzyvisuals.vercel.app">Emmanuel Ariyo</a> Β· Axiveri
</div>
""")
demo.launch()