import gradio as gr
import numpy as np
import os
import time
import torch
import tempfile
import threading
import scipy.io.wavfile
import traceback
from huggingface_hub import login
from pocket_tts import TTSModel
# Configure PyTorch threading behavior for CPU optimization
torch.set_num_threads(1)
torch.set_num_interop_threads(1)
# HF Token for gated models in Spaces
hf_token = os.getenv("HF_TOKEN")
if hf_token:
print("HF_TOKEN found, logging in...")
login(token=hf_token)
VOICES = ['alba', 'marius', 'javert', 'jean', 'fantine', 'cosette', 'eponine', 'azelma']
# Default configuration values
DEFAULT_VOICE = "alba"
DEFAULT_MODEL_VARIANT = "b6369a24"
DEFAULT_TEMPERATURE = 0.1
DEFAULT_LSD_DECODE_STEPS = 1
DEFAULT_EOS_THRESHOLD = -4.0
DEFAULT_NOISE_CLAMP = 0.0
DEFAULT_FRAMES_AFTER_EOS = 10
MAXIMUM_INPUT_LENGTH = 111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111
TEMPORARY_FILE_LIFETIME_SECONDS = 7200 # 2 hours
generation_state_lock = threading.Lock()
is_currently_generating = False
stop_generation_requested = False
temporary_files_registry = {}
temporary_files_lock = threading.Lock()
class TextToSpeechManager:
"""
Manages TTS model lifecycle and speech generation operations.
Implements lazy loading and caching strategies for performance.
"""
def __init__(self):
self.loaded_model = None
self.current_configuration = {}
self.voice_state_cache = {}
def load_or_get_model(
self,
model_variant,
temperature,
lsd_decode_steps,
noise_clamp,
eos_threshold
):
"""Load a TTS model or return cached instance if configuration matches."""
processed_variant = str(model_variant or DEFAULT_MODEL_VARIANT).strip()
processed_temperature = float(temperature) if temperature is not None else DEFAULT_TEMPERATURE
processed_lsd_steps = int(lsd_decode_steps) if lsd_decode_steps is not None else DEFAULT_LSD_DECODE_STEPS
processed_noise_clamp = float(noise_clamp) if noise_clamp and float(noise_clamp) > 0 else None
processed_eos_threshold = float(eos_threshold) if eos_threshold is not None else DEFAULT_EOS_THRESHOLD
requested_configuration = {
"variant": processed_variant,
"temp": processed_temperature,
"lsd_decode_steps": processed_lsd_steps,
"noise_clamp": processed_noise_clamp,
"eos_threshold": processed_eos_threshold
}
if self.loaded_model is None or self.current_configuration != requested_configuration:
print(f"Loading model with config: {requested_configuration}")
self.loaded_model = TTSModel.load_model(**requested_configuration)
self.current_configuration = requested_configuration
self.voice_state_cache = {}
print("Model loaded.")
return self.loaded_model
def get_voice_state_for_preset(self, voice_name):
"""Get or compute voice state for a preset voice with caching."""
validated_voice = voice_name if voice_name in VOICES else DEFAULT_VOICE
if validated_voice not in self.voice_state_cache:
self.voice_state_cache[validated_voice] = self.loaded_model.get_state_for_audio_prompt(
audio_conditioning=validated_voice,
truncate=False
)
return self.voice_state_cache[validated_voice]
def get_voice_state_for_clone(self, audio_file_path):
"""Compute voice state from uploaded audio file for voice cloning."""
return self.loaded_model.get_state_for_audio_prompt(
audio_conditioning=audio_file_path,
truncate=False
)
def generate_audio(self, text_content, voice_state, frames_after_eos, enable_custom_frames):
"""Generate speech audio from text using the specified voice state."""
processed_frames = int(frames_after_eos) if enable_custom_frames else None
return self.loaded_model.generate_audio(
model_state=voice_state,
text_to_generate=text_content,
frames_after_eos=processed_frames,
copy_state=True
)
def save_audio_to_file(self, audio_tensor):
"""Save generated audio tensor to a temporary WAV file."""
audio_numpy_data = audio_tensor.numpy()
audio_sample_rate = self.loaded_model.sample_rate
output_file = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
scipy.io.wavfile.write(output_file.name, audio_sample_rate, audio_numpy_data)
with temporary_files_lock:
temporary_files_registry[output_file.name] = time.time()
return output_file.name
# Create global TTS manager instance
tts_manager = TextToSpeechManager()
# Load model at startup with default parameters
print("Loading PocketTTS model with default parameters...")
tts_manager.load_or_get_model(
DEFAULT_MODEL_VARIANT,
DEFAULT_TEMPERATURE,
DEFAULT_LSD_DECODE_STEPS,
DEFAULT_NOISE_CLAMP,
DEFAULT_EOS_THRESHOLD
)
print("Model ready!")
def cleanup_expired_temporary_files():
"""Remove temporary files that have exceeded their lifetime."""
current_timestamp = time.time()
expired_files = []
with temporary_files_lock:
for file_path, creation_timestamp in list(temporary_files_registry.items()):
if current_timestamp - creation_timestamp > TEMPORARY_FILE_LIFETIME_SECONDS:
expired_files.append(file_path)
for file_path in expired_files:
try:
if os.path.exists(file_path):
os.remove(file_path)
del temporary_files_registry[file_path]
except Exception:
pass
def validate_text_input(text_content):
"""Validate and clean text input for speech generation."""
if not text_content or not isinstance(text_content, str):
return False, ""
cleaned_text = text_content.strip()
if not cleaned_text:
return False, ""
if len(cleaned_text) > MAXIMUM_INPUT_LENGTH:
return False, f"Input exceeds maximum length of {MAXIMUM_INPUT_LENGTH} characters."
return True, cleaned_text
def request_generation_stop():
"""Signal a request to stop the current generation."""
global stop_generation_requested
stop_generation_requested = True
return gr.update(interactive=False)
# Speech generation function
def generate_speech(
text,
voice_mode,
voice_dropdown,
voice_upload,
temperature,
lsd_decode_steps,
noise_clamp,
eos_threshold,
frames_after_eos,
enable_custom_frames
):
"""Perform the complete speech generation workflow with thread safety."""
global is_currently_generating, stop_generation_requested
cleanup_expired_temporary_files()
is_valid, validation_result = validate_text_input(text)
if not is_valid:
if validation_result:
raise gr.Error(validation_result)
raise gr.Error("Please enter valid text to generate speech.")
if voice_mode == "Voice Cloning" and not voice_upload:
raise gr.Error("Please upload an audio file for voice cloning.")
with generation_state_lock:
if is_currently_generating:
raise gr.Error("A generation is already in progress. Please wait.")
is_currently_generating = True
stop_generation_requested = False
try:
tts_manager.load_or_get_model(
DEFAULT_MODEL_VARIANT,
temperature,
lsd_decode_steps,
noise_clamp,
eos_threshold
)
if stop_generation_requested:
return None
if voice_mode == "Voice Cloning":
voice_state = tts_manager.get_voice_state_for_clone(voice_upload)
else:
voice_state = tts_manager.get_voice_state_for_preset(voice_dropdown)
if stop_generation_requested:
return None
print(f"Generating with voice mode: {voice_mode}, temp: {temperature}, lsd_steps: {lsd_decode_steps}")
generated_audio = tts_manager.generate_audio(
validation_result,
voice_state,
frames_after_eos,
enable_custom_frames
)
if stop_generation_requested:
return None
output_file_path = tts_manager.save_audio_to_file(generated_audio)
return output_file_path
except gr.Error:
raise
except Exception as e:
full_error = traceback.format_exc()
print(f"Unexpected error: {full_error}")
raise gr.Error(f"An unexpected error occurred: {str(e)}")
finally:
with generation_state_lock:
is_currently_generating = False
stop_generation_requested = False
# Load custom theme with fallback
try:
theme = gr.Theme.from_hub("JohnSmith9982/small_and_pretty")
except Exception as e:
print(f"Warning: Could not load custom theme: {e}. Using default Soft theme.")
theme = gr.themes.Soft()
css = """
footer {visibility: hidden}
.gradio-container {
max-width: 100% !important;
padding: 0 !important;
}
@media (min-width: 768px) {
.gradio-container {
padding-left: 2% !important;
padding-right: 2% !important;
}
}
.header-section {
text-align: left;
margin-bottom: 1.5rem;
}
.main-title {
color: #10b981;
font-weight: 800;
font-size: 1.8rem;
margin: 5px 0;
}
@media (min-width: 768px) {
.main-title {
font-size: 2.2rem;
}
}
.logo-container {
display: flex;
justify-content: flex-start;
align-items: center;
gap: 10px;
margin-bottom: 10px;
}
.logo-img {
height: 40px;
border-radius: 8px;
}
@media (min-width: 768px) {
.logo-img {
height: 50px;
}
.logo-container {
gap: 15px;
}
}
.description {
max-width: 900px;
margin: 10px 0;
font-size: 0.95rem;
line-height: 1.5;
color: #4b5563;
}
.links-row {
display: flex;
flex-wrap: wrap;
justify-content: flex-start;
gap: 8px;
margin: 10px 0;
font-size: 0.85rem;
}
@media (min-width: 768px) {
.links-row {
gap: 10px;
font-size: 0.9rem;
}
}
.links-row a {
color: #10b981;
text-decoration: none;
padding: 3px 10px;
border: 1px solid #10b981;
border-radius: 15px;
transition: all 0.2s;
white-space: nowrap;
}
.links-row a:hover {
background-color: #10b981;
color: white;
}
.social-handles {
display: flex;
justify-content: center;
gap: 20px;
margin: 15px 0;
}
.social-icon {
width: 28px;
height: 28px;
transition: all 0.3s ease;
}
.social-icon:hover {
transform: scale(1.1) translateY(-3px);
}
.disclaimer {
text-align: center;
font-size: 0.8rem;
color: #9ca3af;
margin-top: 30px;
padding: 15px;
border-top: 1px solid #f3f4f6;
}
@media (min-width: 768px) {
.disclaimer {
margin-top: 40px;
padding: 20px;
}
}
#voice-mode .wrap {
display: flex !important;
flex-direction: row !important;
width: 100% !important;
}
#voice-mode .wrap label {
flex: 1 !important;
justify-content: center !important;
text-align: center !important;
}
"""
with gr.Blocks() as demo:
with gr.Column(elem_classes="header-section"):
with gr.Row():
with gr.Column(scale=4):
gr.HTML("""
XlnkTTS
""")
gr.HTML("""
""")
gr.HTML("""
""")
with gr.Row():
with gr.Column(scale=1):
text_input = gr.Textbox(
label="Text to Speak",
placeholder="Enter text here...",
value="Hello! Welcome to Pocket TTS. This lightweight text to speech model runs entirely on your CPU. Try changing the voice or adjusting the generation parameters below.",
lines=9,
elem_id="text-input"
)
voice_mode = gr.Radio(
choices=["Kyutai Voices", "Voice Cloning"],
value="Kyutai Voices",
label="Voice Mode",
elem_id="voice-mode"
)
with gr.Column(visible=True) as standard_voice_col:
voice_select = gr.Dropdown(
choices=VOICES,
value="alba",
label="Select from Kyutai Voices",
elem_id="voice-select"
)
with gr.Column(visible=False) as cloning_voice_col:
voice_upload = gr.Audio(
label="Upload Voice for Cloning (WAV/MP3)",
type="filepath",
elem_id="voice-upload"
)
# Generation Parameters Accordion
with gr.Accordion("⚙️ Generation Parameters", open=False):
with gr.Row():
temperature_slider = gr.Slider(
label="Temperature",
minimum=0.1,
maximum=2.0,
step=0.05,
value=DEFAULT_TEMPERATURE,
info="Higher values produce more expressive speech"
)
lsd_decode_steps_slider = gr.Slider(
label="LSD Decode Steps",
minimum=1,
maximum=20,
step=1,
value=DEFAULT_LSD_DECODE_STEPS,
info="More steps may improve quality but slower"
)
with gr.Row():
noise_clamp_slider = gr.Slider(
label="Noise Clamp",
minimum=0.0,
maximum=2.0,
step=0.05,
value=DEFAULT_NOISE_CLAMP,
info="Maximum noise sampling value (0 = disabled)"
)
eos_threshold_slider = gr.Slider(
label="End of Sequence Threshold",
minimum=-10.0,
maximum=0.0,
step=0.25,
value=DEFAULT_EOS_THRESHOLD,
info="Smaller values cause earlier completion"
)
with gr.Row():
enable_custom_frames_checkbox = gr.Checkbox(
label="Enable Custom Frames After EOS",
value=False,
info="Manually control post-EOS frame generation"
)
frames_after_eos_slider = gr.Slider(
label="Frames After EOS",
minimum=0,
maximum=100,
step=1,
value=DEFAULT_FRAMES_AFTER_EOS,
info="Additional frames after end-of-sequence (80ms per frame)"
)
with gr.Row():
clear_btn = gr.Button("🗑️ Clear", variant="secondary")
generate_btn = gr.Button("⚡ Generate", variant="primary")
stop_btn = gr.Button("🔴 Stop", variant="stop", visible=False)
with gr.Column(scale=1):
audio_output = gr.Audio(
label="Audio Output",
autoplay=True,
elem_id="audio-output"
)
gr.Examples(
examples=[
["On Tuesday, the seventeenth of October, two thousand twenty-five, at exactly six forty-five in the morning, the outdoor temperature dropped to twelve point eight degrees Celsius. The forecast predicts a high of twenty-two degrees by noon.", "alba"],
["Welcome to Station Forty-Seven. Your train to Platform Nineteen B will arrive in approximately fifteen minutes. Please have your tickets ready for inspection.", "marius"],
["You dare defy me? I have spent twenty long years hunting you across every shadow and every corner of this wretched kingdom. There is no escape. Justice will find you, and when it does, you will kneel before me and beg for mercy that will never come!", "javert"],
["Flight Seven Ninety-Two to London Heathrow is now boarding at Gate Twenty-Three A. Final call for passengers Smith and Johnson. Departure is scheduled for fourteen thirty hours.", "jean"],
["Our quarterly revenue reached four point seven million dollars, up eighteen percent from last year. The board meeting is scheduled for the twenty-fifth of November at two fifteen in the afternoon.", "fantine"],
["The recipe calls for three hundred fifty grams of flour, two hundred milliliters of milk, and one point five teaspoons of vanilla extract. Bake at one hundred eighty degrees for forty-five minutes.", "cosette"],
["Chapter Fourteen, Page Two Hundred Thirty-Seven. The mysterious traveler arrived at the inn precisely at midnight. He carried nothing but a worn leather satchel and spoke with an accent no one could place.", "eponine"],
["Exercise routine: Run five kilometers in under thirty minutes. Complete three sets of fifteen push-ups. Rest for ninety seconds between each set. Cool down with ten minutes of stretching.", "azelma"]
],
inputs=[text_input, voice_select],
)
gr.HTML("""
""")
# Visibility Toggling
def update_voice_ui(mode):
if mode == "Kyutai Voices":
return gr.update(visible=True), gr.update(visible=False)
else:
return gr.update(visible=False), gr.update(visible=True)
voice_mode.change(
fn=update_voice_ui,
inputs=[voice_mode],
outputs=[standard_voice_col, cloning_voice_col]
)
# Define generation inputs list
generation_inputs = [
text_input,
voice_mode,
voice_select,
voice_upload,
temperature_slider,
lsd_decode_steps_slider,
noise_clamp_slider,
eos_threshold_slider,
frames_after_eos_slider,
enable_custom_frames_checkbox
]
# UI state management functions
def switch_to_generating_state():
return (
gr.update(visible=False), # Hide generate button
gr.update(visible=True, interactive=True) # Show stop button
)
def switch_to_idle_state():
return (
gr.update(visible=True), # Show generate button
gr.update(visible=False) # Hide stop button
)
# Event handlers
generate_btn.click(
fn=switch_to_generating_state,
outputs=[generate_btn, stop_btn]
).then(
fn=generate_speech,
inputs=generation_inputs,
outputs=audio_output
).then(
fn=switch_to_idle_state,
outputs=[generate_btn, stop_btn]
)
text_input.submit(
fn=switch_to_generating_state,
outputs=[generate_btn, stop_btn]
).then(
fn=generate_speech,
inputs=generation_inputs,
outputs=audio_output
).then(
fn=switch_to_idle_state,
outputs=[generate_btn, stop_btn]
)
# Stop button handler
stop_btn.click(
fn=request_generation_stop,
outputs=[stop_btn]
)
# Clear button handler - also reset generation parameters
def perform_clear_action():
return (
"", # text_input
"Kyutai Voices", # voice_mode
"alba", # voice_select
None, # voice_upload
None, # audio_output
DEFAULT_TEMPERATURE, # temperature_slider
DEFAULT_LSD_DECODE_STEPS, # lsd_decode_steps_slider
DEFAULT_NOISE_CLAMP, # noise_clamp_slider
DEFAULT_EOS_THRESHOLD, # eos_threshold_slider
DEFAULT_FRAMES_AFTER_EOS, # frames_after_eos_slider
False # enable_custom_frames_checkbox
)
clear_btn.click(
fn=perform_clear_action,
outputs=[
text_input,
voice_mode,
voice_select,
voice_upload,
audio_output,
temperature_slider,
lsd_decode_steps_slider,
noise_clamp_slider,
eos_threshold_slider,
frames_after_eos_slider,
enable_custom_frames_checkbox
]
)
if __name__ == "__main__":
demo.queue().launch(theme=theme, css=css)