Update app.py
Browse files
app.py
CHANGED
@@ -3,7 +3,7 @@ import torchaudio
|
|
3 |
import gradio as gr
|
4 |
|
5 |
from zonos.model import Zonos
|
6 |
-
from zonos.conditioning import make_cond_dict
|
7 |
|
8 |
# Global cache to hold the loaded model
|
9 |
MODEL = None
|
@@ -12,7 +12,7 @@ device = "cuda"
|
|
12 |
def load_model():
|
13 |
"""
|
14 |
Loads the Zonos model once and caches it globally.
|
15 |
-
Adjust the model name
|
16 |
"""
|
17 |
global MODEL
|
18 |
if MODEL is None:
|
@@ -20,26 +20,29 @@ def load_model():
|
|
20 |
print(f"Loading model: {model_name}")
|
21 |
MODEL = Zonos.from_pretrained(model_name, device="cuda")
|
22 |
MODEL = MODEL.requires_grad_(False).eval()
|
23 |
-
MODEL.bfloat16() # optional
|
24 |
print("Model loaded successfully!")
|
25 |
return MODEL
|
26 |
|
27 |
-
def tts(text, speaker_audio):
|
28 |
"""
|
29 |
text: str
|
30 |
speaker_audio: (sample_rate, numpy_array) from Gradio if type="numpy"
|
|
|
|
|
31 |
Returns (sample_rate, waveform) for Gradio audio output.
|
32 |
"""
|
33 |
model = load_model()
|
34 |
|
|
|
35 |
if not text:
|
36 |
return None
|
37 |
|
38 |
-
# If
|
39 |
if speaker_audio is None:
|
40 |
return None
|
41 |
|
42 |
-
# Gradio provides audio in
|
43 |
sr, wav_np = speaker_audio
|
44 |
|
45 |
# Convert to Torch tensor: shape (1, num_samples)
|
@@ -55,17 +58,15 @@ def tts(text, speaker_audio):
|
|
55 |
|
56 |
# Prepare conditioning dictionary
|
57 |
cond_dict = make_cond_dict(
|
58 |
-
text=text,
|
59 |
-
speaker=spk_embedding,
|
60 |
-
language=
|
61 |
device=device,
|
62 |
)
|
63 |
conditioning = model.prepare_conditioning(cond_dict)
|
64 |
|
65 |
# Generate codes
|
66 |
with torch.no_grad():
|
67 |
-
# Optionally set a manual seed for reproducibility
|
68 |
-
# torch.manual_seed(1234)
|
69 |
codes = model.generate(conditioning)
|
70 |
|
71 |
# Decode the codes into raw audio
|
@@ -76,7 +77,7 @@ def tts(text, speaker_audio):
|
|
76 |
|
77 |
def build_demo():
|
78 |
with gr.Blocks() as demo:
|
79 |
-
gr.Markdown("# Simple Zonos TTS Demo (Text + Reference Audio)")
|
80 |
|
81 |
with gr.Row():
|
82 |
text_input = gr.Textbox(
|
@@ -88,16 +89,26 @@ def build_demo():
|
|
88 |
label="Reference Audio (Speaker Cloning)",
|
89 |
type="numpy"
|
90 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
91 |
|
92 |
generate_button = gr.Button("Generate")
|
93 |
|
94 |
-
# The output
|
95 |
audio_output = gr.Audio(label="Synthesized Output", type="numpy")
|
96 |
|
97 |
-
# Bind the generate button
|
98 |
generate_button.click(
|
99 |
fn=tts,
|
100 |
-
inputs=[text_input, ref_audio_input],
|
101 |
outputs=audio_output,
|
102 |
)
|
103 |
|
|
|
3 |
import gradio as gr
|
4 |
|
5 |
from zonos.model import Zonos
|
6 |
+
from zonos.conditioning import make_cond_dict, supported_language_codes
|
7 |
|
8 |
# Global cache to hold the loaded model
|
9 |
MODEL = None
|
|
|
12 |
def load_model():
|
13 |
"""
|
14 |
Loads the Zonos model once and caches it globally.
|
15 |
+
Adjust the model name if you want to switch from hybrid to transformer, etc.
|
16 |
"""
|
17 |
global MODEL
|
18 |
if MODEL is None:
|
|
|
20 |
print(f"Loading model: {model_name}")
|
21 |
MODEL = Zonos.from_pretrained(model_name, device="cuda")
|
22 |
MODEL = MODEL.requires_grad_(False).eval()
|
23 |
+
MODEL.bfloat16() # optional if your GPU supports bfloat16
|
24 |
print("Model loaded successfully!")
|
25 |
return MODEL
|
26 |
|
27 |
+
def tts(text, speaker_audio, selected_language):
|
28 |
"""
|
29 |
text: str
|
30 |
speaker_audio: (sample_rate, numpy_array) from Gradio if type="numpy"
|
31 |
+
selected_language: str (e.g., "en-us", "es-es", etc.)
|
32 |
+
|
33 |
Returns (sample_rate, waveform) for Gradio audio output.
|
34 |
"""
|
35 |
model = load_model()
|
36 |
|
37 |
+
# If no text, return None
|
38 |
if not text:
|
39 |
return None
|
40 |
|
41 |
+
# If no reference audio, return None
|
42 |
if speaker_audio is None:
|
43 |
return None
|
44 |
|
45 |
+
# Gradio provides audio in (sample_rate, numpy_array)
|
46 |
sr, wav_np = speaker_audio
|
47 |
|
48 |
# Convert to Torch tensor: shape (1, num_samples)
|
|
|
58 |
|
59 |
# Prepare conditioning dictionary
|
60 |
cond_dict = make_cond_dict(
|
61 |
+
text=text, # The text prompt
|
62 |
+
speaker=spk_embedding, # Speaker embedding
|
63 |
+
language=selected_language, # Language from the Dropdown
|
64 |
device=device,
|
65 |
)
|
66 |
conditioning = model.prepare_conditioning(cond_dict)
|
67 |
|
68 |
# Generate codes
|
69 |
with torch.no_grad():
|
|
|
|
|
70 |
codes = model.generate(conditioning)
|
71 |
|
72 |
# Decode the codes into raw audio
|
|
|
77 |
|
78 |
def build_demo():
|
79 |
with gr.Blocks() as demo:
|
80 |
+
gr.Markdown("# Simple Zonos TTS Demo (Text + Reference Audio + Language)")
|
81 |
|
82 |
with gr.Row():
|
83 |
text_input = gr.Textbox(
|
|
|
89 |
label="Reference Audio (Speaker Cloning)",
|
90 |
type="numpy"
|
91 |
)
|
92 |
+
# Add a dropdown for language selection
|
93 |
+
language_dropdown = gr.Dropdown(
|
94 |
+
label="Language",
|
95 |
+
# You can provide your own subset or use all:
|
96 |
+
# For demonstration, let's pick 5 common ones
|
97 |
+
# or you can do: choices=supported_language_codes
|
98 |
+
choices=["en-us", "es-es", "fr-fr", "de-de", "it"],
|
99 |
+
value="en-us",
|
100 |
+
interactive=True
|
101 |
+
)
|
102 |
|
103 |
generate_button = gr.Button("Generate")
|
104 |
|
105 |
+
# The output is an audio widget that Gradio will play
|
106 |
audio_output = gr.Audio(label="Synthesized Output", type="numpy")
|
107 |
|
108 |
+
# Bind the generate button: pass text, reference audio, and selected language
|
109 |
generate_button.click(
|
110 |
fn=tts,
|
111 |
+
inputs=[text_input, ref_audio_input, language_dropdown],
|
112 |
outputs=audio_output,
|
113 |
)
|
114 |
|