Kokoro-TTS-Zero / tts_model.py
Remsky's picture
Added Multi-Voice, GPU Timeout, etc
3c8cbc9
import os
import torch
import numpy as np
import time
import matplotlib.pyplot as plt
from typing import Tuple, List
from statistics import mean, median, stdev
from lib import (
normalize_text,
chunk_text,
count_tokens,
load_module_from_file,
download_model_files,
list_voice_files,
download_voice_files,
ensure_dir,
concatenate_audio_chunks
)
import spaces
class TTSModel:
"""GPU-accelerated TTS model manager"""
def __init__(self):
self.model = None
self.voices_dir = "voices"
self.model_repo = "hexgrad/Kokoro-82M"
ensure_dir(self.voices_dir)
self.model_path = None
# Load required modules
py_modules = ["istftnet", "plbert", "models", "kokoro"]
module_files = download_model_files(self.model_repo, [f"{m}.py" for m in py_modules])
for module_name, file_path in zip(py_modules, module_files):
load_module_from_file(module_name, file_path)
# Import required functions from kokoro module
kokoro = __import__("kokoro")
self.generate = kokoro.generate
self.build_model = __import__("models").build_model
def initialize(self) -> bool:
"""Initialize model and download voices"""
try:
print("Initializing model...")
# Download model files
model_files = download_model_files(
self.model_repo,
["kokoro-v0_19.pth", "config.json"]
)
self.model_path = model_files[0] # kokoro-v0_19.pth
# Download voice files
download_voice_files(self.model_repo, "voices", self.voices_dir)
# Get list of available voices
available_voices = self.list_voices()
print("Model initialization complete")
return True
except Exception as e:
print(f"Error initializing model: {str(e)}")
return False
def ensure_voice_downloaded(self, voice_name: str) -> bool:
"""Ensure specific voice is downloaded"""
try:
voice_path = os.path.join(self.voices_dir, "voices", f"{voice_name}.pt")
if not os.path.exists(voice_path):
print(f"Downloading voice {voice_name}.pt...")
download_voice_files(self.model_repo, [f"{voice_name}.pt"], self.voices_dir)
return True
except Exception as e:
print(f"Error downloading voice {voice_name}: {str(e)}")
return False
def list_voices(self) -> List[str]:
"""List available voices"""
voices = []
voices_subdir = os.path.join(self.voices_dir, "voices")
if os.path.exists(voices_subdir):
for file in os.listdir(voices_subdir):
if file.endswith(".pt"):
voice_name = file[:-3]
voices.append(voice_name)
return voices
# def _ensure_model_on_gpu(self) -> None:
# """Ensure model is on GPU and stays there"""
# if not hasattr(self, '_model_on_gpu') or not self._model_on_gpu:
# print("Moving model to GPU...")
# with torch.cuda.device(0):
# torch.cuda.set_device(0)
# if hasattr(self.model, 'to'):
# self.model.to('cuda')
# else:
# for name in self.model:
# if isinstance(self.model[name], torch.Tensor):
# self.model[name] = self.model[name].cuda()
# self._model_on_gpu = True
def _generate_audio(self, text: str, voicepack: torch.Tensor, lang: str, speed: float) -> np.ndarray:
"""GPU-accelerated audio generation"""
try:
with torch.cuda.device(0):
torch.cuda.set_device(0)
try:
# Build model if needed
if self.model is None:
print("Building model...")
device = torch.device('cuda')
self.model = self.build_model(self.model_path, device=device)
if self.model is None:
raise ValueError("Failed to build model")
print("Model built successfully")
# Move model to GPU if needed
if not hasattr(self.model, '_on_gpu'):
print("Moving model to GPU...")
if hasattr(self.model, 'to'):
self.model = self.model.to('cuda')
else:
for name in self.model:
if isinstance(self.model[name], torch.Tensor):
self.model[name] = self.model[name].cuda()
self.model._on_gpu = True
except Exception as e:
print(f"Error building model: {str(e)}")
print("Attempting to continue")
raise e
# Move voicepack to GPU
voicepack = voicepack.cuda()
# Run generation with everything on GPU
audio, _ = self.generate(
self.model,
text,
voicepack,
lang=lang,
speed=speed
)
return audio
except Exception as e:
print(f"Error in audio generation: {str(e)}")
raise e
@spaces.GPU(duration=None) # Duration will be set by the UI
def generate_speech(self, text: str, voice_names: list[str], speed: float = 1.0, gpu_timeout: int = 60, progress_callback=None, progress_state=None, progress=None) -> Tuple[np.ndarray, float]:
"""Generate speech from text. Returns (audio_array, duration)
Args:
text: Input text to convert to speech
voice_name: Name of voice to use
speed: Speech speed multiplier
progress_callback: Optional callback function(chunk_num, total_chunks, tokens_per_sec, rtf, progress_state, start_time, gpu_timeout, progress)
progress_state: Dictionary tracking generation progress metrics
progress: Progress callback from Gradio
"""
try:
start_time = time.time()
with torch.cuda.device(0):
torch.cuda.set_device(0)
if not text or not voice_names:
raise ValueError("Text and voice name are required")
# Build model directly on GPU
# Build model if needed
if self.model is None:
print("Building model...")
self.model = self.build_model(self.model_path, device='cuda')
if self.model is None:
raise ValueError("Failed to build model")
print("Model built successfully")
# Move model to GPU if needed
if not hasattr(self.model, '_on_gpu'):
print("Moving model to GPU...")
if hasattr(self.model, 'to'):
self.model = self.model.to('cuda')
else:
for name in self.model:
if isinstance(self.model[name], torch.Tensor):
self.model[name] = self.model[name].cuda()
self.model._on_gpu = True
t_voices = []
if isinstance(voice_names, list) and len(voice_names) > 1:
for voice in voice_names:
try:
voice_path = os.path.join(self.voices_dir, "voices", f"{voice}.pt")
voicepack = torch.load(voice_path, weights_only=True)
t_voices.append(voicepack)
except Exception as e:
print(f"Warning: Failed to load voice {voice}: {str(e)}")
# Combine voices by taking mean
voicepack = torch.mean(torch.stack(t_voices), dim=0)
voice_name = "_".join(voice_names)
else:
voice_name = voice_names[0]
voice_path = os.path.join(self.voices_dir, "voices", f"{voice_name}.pt")
voicepack = torch.load(voice_path, weights_only=True)
# Count tokens and normalize text
total_tokens = count_tokens(text)
text = normalize_text(text)
if not text:
raise ValueError("Text is empty after normalization")
# Break text into chunks for better memory management
chunks = chunk_text(text)
print(f"Processing {len(chunks)} chunks...")
# Process all chunks within same GPU context
audio_chunks = []
chunk_times = []
chunk_sizes = [] # Store chunk lengths
total_processed_tokens = 0
total_processed_time = 0
for i, chunk in enumerate(chunks):
chunk_start = time.time()
chunk_audio = self._generate_audio(
text=chunk,
voicepack=voicepack,
lang=voice_name[0],
speed=speed
)
chunk_time = time.time() - chunk_start
# Calculate per-chunk metrics
chunk_tokens = count_tokens(chunk)
chunk_tokens_per_sec = chunk_tokens / chunk_time
# Update totals for overall stats
total_processed_tokens += chunk_tokens
total_processed_time += chunk_time
# Calculate processing speed metrics
chunk_duration = len(chunk_audio) / 24000 # audio duration in seconds
rtf = chunk_time / chunk_duration
times_faster = 1 / rtf
chunk_times.append(chunk_time)
chunk_sizes.append(len(chunk))
print(f"Chunk {i+1}/{len(chunks)} processed in {chunk_time:.2f}s")
print(f"Current tokens/sec: {chunk_tokens_per_sec:.2f}")
print(f"Real-time factor: {rtf:.2f}x")
print(f"{times_faster:.1f}x faster than real-time")
audio_chunks.append(chunk_audio)
# Call progress callback if provided
if progress_callback:
progress_callback(
i + 1, # chunk_num
len(chunks), # total_chunks
chunk_tokens_per_sec, # Pass per-chunk rate instead of cumulative
rtf,
progress_state, # Added
start_time, # Added
gpu_timeout, # Use the timeout value from UI
progress # Added
)
# Concatenate audio chunks
audio = concatenate_audio_chunks(audio_chunks)
# Return audio and metrics
return (
audio, # Audio array
len(audio) / 24000, # Duration
{
"chunk_times": chunk_times,
"chunk_sizes": chunk_sizes,
"tokens_per_sec": [float(x) for x in progress_state["tokens_per_sec"]],
"rtf": [float(x) for x in progress_state["rtf"]],
"total_tokens": total_tokens,
"total_time": time.time() - start_time
}
)
except Exception as e:
print(f"Error generating speech: {str(e)}")
raise