Spaces:
Runtime error
Runtime error
initial setup
Browse files- app.py +165 -0
- nanospeech/voices/celeste.txt +1 -0
- nanospeech/voices/celeste.wav +0 -0
- nanospeech/voices/luna.txt +1 -0
- nanospeech/voices/luna.wav +0 -0
- nanospeech/voices/nash.txt +1 -0
- nanospeech/voices/nash.wav +0 -0
- nanospeech/voices/orion.txt +1 -0
- nanospeech/voices/orion.wav +0 -0
- nanospeech/voices/rhea.txt +1 -0
- nanospeech/voices/rhea.wav +0 -0
- requirements.txt +4 -0
app.py
ADDED
@@ -0,0 +1,165 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from glob import glob
|
3 |
+
from dataclasses import dataclass
|
4 |
+
|
5 |
+
import gradio as gr
|
6 |
+
import soundfile as sf
|
7 |
+
from nanospeech.nanospeech_torch import Nanospeech
|
8 |
+
from nanospeech.generate import generate_one, SAMPLE_RATE, split_sentences
|
9 |
+
import numpy as np
|
10 |
+
from typing import Optional
|
11 |
+
|
12 |
+
|
13 |
+
PROMPT_DIR = 'nanospeech/voices'
|
14 |
+
|
15 |
+
# Note: gradio expects audio as int16, so we need to convert to float32 when loading and convert back when returning
|
16 |
+
|
17 |
+
def convert_audio_int16_to_float32(audio: np.ndarray) -> np.ndarray:
|
18 |
+
return audio.astype(np.float32) / 32768.0
|
19 |
+
|
20 |
+
def convert_audio_float32_to_int16(audio: np.ndarray) -> np.ndarray:
|
21 |
+
return (np.clip(audio, -1.0, 1.0) * 32768.0).astype(np.int16)
|
22 |
+
|
23 |
+
@dataclass
|
24 |
+
class VoicePrompt:
|
25 |
+
wav_path: str
|
26 |
+
text: str
|
27 |
+
|
28 |
+
def get_prompt_list(prompt_dir=PROMPT_DIR):
|
29 |
+
wav_paths = glob(os.path.join(prompt_dir, '*.wav'))
|
30 |
+
|
31 |
+
prompt_lookup: dict[str, VoicePrompt] = {}
|
32 |
+
|
33 |
+
for wav_path in wav_paths:
|
34 |
+
voice_name = os.path.splitext(os.path.basename(wav_path))[0]
|
35 |
+
text_path = wav_path.replace('.wav', '.txt')
|
36 |
+
|
37 |
+
with open(text_path, 'r') as f:
|
38 |
+
text = f.read()
|
39 |
+
|
40 |
+
prompt_lookup[voice_name] = VoicePrompt(
|
41 |
+
wav_path=wav_path,
|
42 |
+
text=text
|
43 |
+
)
|
44 |
+
|
45 |
+
return prompt_lookup
|
46 |
+
|
47 |
+
def create_demo(prompt_list: dict[str, VoicePrompt], model: 'Nanospeech'):
|
48 |
+
|
49 |
+
def update_prompt(voice_name: str):
|
50 |
+
return (
|
51 |
+
prompt_list[voice_name].wav_path,
|
52 |
+
prompt_list[voice_name].text
|
53 |
+
)
|
54 |
+
|
55 |
+
def _generate(prompt_audio: str, prompt_text: str, input_text: str, nfe_steps: int = 8, method: str = "rk4", cfg_strength: float = 2.0, sway_sampling_coef: float = -1.0, speed: float = 1.0, seed: Optional[int] = None):
|
56 |
+
|
57 |
+
print(f'generating: {input_text}, prompt: {prompt_text}, prompt_audio: {prompt_audio}')
|
58 |
+
|
59 |
+
|
60 |
+
# Load reference audio into memory
|
61 |
+
if isinstance(prompt_audio, tuple):
|
62 |
+
sr, ref_audio = prompt_audio
|
63 |
+
|
64 |
+
ref_audio = convert_audio_int16_to_float32(ref_audio)
|
65 |
+
else:
|
66 |
+
ref_audio, sr = sf.read(prompt_audio)
|
67 |
+
print('loaded from path')
|
68 |
+
|
69 |
+
if sr != SAMPLE_RATE:
|
70 |
+
raise ValueError("Reference audio must be mono with a sample rate of 24kHz")
|
71 |
+
|
72 |
+
# Split input text into sentences
|
73 |
+
sentences = split_sentences(input_text)
|
74 |
+
is_single_generation = len(sentences) <= 1
|
75 |
+
|
76 |
+
if is_single_generation:
|
77 |
+
wave = generate_one(
|
78 |
+
model=model,
|
79 |
+
text=input_text,
|
80 |
+
ref_audio=ref_audio,
|
81 |
+
ref_audio_text=prompt_text,
|
82 |
+
steps=nfe_steps,
|
83 |
+
method=method,
|
84 |
+
cfg_strength=cfg_strength,
|
85 |
+
sway_sampling_coef=sway_sampling_coef,
|
86 |
+
speed=speed,
|
87 |
+
seed=seed,
|
88 |
+
player=None,
|
89 |
+
)
|
90 |
+
if hasattr(wave, 'numpy'):
|
91 |
+
wave = wave.numpy()
|
92 |
+
else:
|
93 |
+
# Generate multiple sentences and concatenate
|
94 |
+
output = []
|
95 |
+
for sentence_text in sentences:
|
96 |
+
wave = generate_one(
|
97 |
+
model=model,
|
98 |
+
text=sentence_text,
|
99 |
+
ref_audio=ref_audio,
|
100 |
+
ref_audio_text=prompt_text,
|
101 |
+
steps=nfe_steps,
|
102 |
+
method=method,
|
103 |
+
cfg_strength=cfg_strength,
|
104 |
+
sway_sampling_coef=sway_sampling_coef,
|
105 |
+
speed=speed,
|
106 |
+
seed=seed,
|
107 |
+
player=None,
|
108 |
+
)
|
109 |
+
if hasattr(wave, 'numpy'):
|
110 |
+
wave = wave.numpy()
|
111 |
+
output.append(wave)
|
112 |
+
|
113 |
+
wave = np.concatenate(output, axis=0)
|
114 |
+
|
115 |
+
return (SAMPLE_RATE, wave)
|
116 |
+
|
117 |
+
|
118 |
+
with gr.Blocks() as demo:
|
119 |
+
gr.Markdown("# (Unofficial) Nanospeech Demo")
|
120 |
+
gr.Markdown("A simple, hackable text-to-speech system in PyTorch and MLX - [github](https://github.com/lucasnewman/nanospeech)")
|
121 |
+
|
122 |
+
with gr.Group():
|
123 |
+
gr.Markdown("## Select a voice prompt")
|
124 |
+
voice_dropdown = gr.Dropdown(choices=list(prompt_list.keys()), value='celeste', interactive=True, label="Voice")
|
125 |
+
|
126 |
+
with gr.Group():
|
127 |
+
gr.Markdown("## Voice Prompt")
|
128 |
+
with gr.Row():
|
129 |
+
prompt_audio = gr.Audio(label="Audio", value=prompt_list[voice_dropdown.value].wav_path)
|
130 |
+
prompt_text = gr.Textbox(label="Text", value=prompt_list[voice_dropdown.value].text, interactive=False)
|
131 |
+
|
132 |
+
voice_dropdown.change(fn=update_prompt, inputs=voice_dropdown, outputs=[prompt_audio, prompt_text])
|
133 |
+
|
134 |
+
with gr.Accordion("Advanced Settings", open=False):
|
135 |
+
speed = gr.Slider(label="Speed", value=1.0, minimum=0.1, maximum=2.0, step=0.1)
|
136 |
+
nfe_steps = gr.Slider(label="NFE Steps - more steps = more stable, but slower", value=8, minimum=1, maximum=64, step=1)
|
137 |
+
|
138 |
+
method = gr.Dropdown(choices=["rk4", "euler", "midpoint"], value="rk4", label="Method")
|
139 |
+
cfg_strength = gr.Slider(label="CFG Strength", value=2.0, minimum=0.0, maximum=5.0, step=0.1)
|
140 |
+
sway_sampling_coef = gr.Slider(label="Sway Sampling Coef", value=-1.0, minimum=-5.0, maximum=5.0, step=0.1)
|
141 |
+
|
142 |
+
with gr.Group():
|
143 |
+
gr.Markdown("# Generate")
|
144 |
+
|
145 |
+
input_text = gr.Textbox(label="Input Text", value="Hello, how are you?")
|
146 |
+
generate_button = gr.Button("Generate")
|
147 |
+
|
148 |
+
with gr.Group():
|
149 |
+
output_audio = gr.Audio(label="Output Audio")
|
150 |
+
|
151 |
+
generate_button.click(fn=_generate, inputs=[prompt_audio, prompt_text, input_text, nfe_steps, method, cfg_strength, sway_sampling_coef, speed], outputs=output_audio)
|
152 |
+
|
153 |
+
return demo
|
154 |
+
|
155 |
+
|
156 |
+
if __name__ == "__main__":
|
157 |
+
|
158 |
+
|
159 |
+
# Preload the model
|
160 |
+
model = Nanospeech.from_pretrained("lucasnewman/nanospeech")
|
161 |
+
prompt_list = get_prompt_list()
|
162 |
+
|
163 |
+
demo = create_demo(prompt_list, model)
|
164 |
+
demo.launch()
|
165 |
+
|
nanospeech/voices/celeste.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
Pickled cucumbers.
|
nanospeech/voices/celeste.wav
ADDED
Binary file (53.8 kB). View file
|
|
nanospeech/voices/luna.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
Do you wish to see him?
|
nanospeech/voices/luna.wav
ADDED
Binary file (50 kB). View file
|
|
nanospeech/voices/nash.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
Look out for what?
|
nanospeech/voices/nash.wav
ADDED
Binary file (50 kB). View file
|
|
nanospeech/voices/orion.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
It isn't his fault.
|
nanospeech/voices/orion.wav
ADDED
Binary file (48 kB). View file
|
|
nanospeech/voices/rhea.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
They haven't met?
|
nanospeech/voices/rhea.wav
ADDED
Binary file (51.9 kB). View file
|
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
gradio
|
2 |
+
soundfile
|
3 |
+
git+https://github.com/thunn/nanospeech.git
|
4 |
+
numpy
|