VoiceAPI / src /engine.py
Harshil748's picture
Initial HF Spaces deployment - downloads models at runtime
ecde958
raw
history blame
19.8 kB
"""
Main TTS Engine for SYSPIN Multi-lingual TTS
Loads and runs VITS models for inference
Supports:
- JIT traced models (.pt) - Hindi, Bengali, Kannada, etc.
- Coqui TTS checkpoints (.pth) - Bhojpuri, etc.
- Facebook MMS models - Gujarati
Includes style/prosody control
"""
import os
import logging
from pathlib import Path
from typing import Dict, Optional, Union, List, Tuple, Any
import numpy as np
import torch
from dataclasses import dataclass
from .config import LANGUAGE_CONFIGS, LanguageConfig, MODELS_DIR, STYLE_PRESETS
from .tokenizer import TTSTokenizer, CharactersConfig, TextNormalizer
from .downloader import ModelDownloader
logger = logging.getLogger(__name__)
logger = logging.getLogger(__name__)
@dataclass
class TTSOutput:
"""Output from TTS synthesis"""
audio: np.ndarray
sample_rate: int
duration: float
voice: str
text: str
style: Optional[str] = None
class StyleProcessor:
"""
Simple prosody/style control via audio post-processing
Supports pitch shifting, speed change, and energy modification
"""
@staticmethod
def apply_pitch_shift(
audio: np.ndarray, sample_rate: int, pitch_factor: float
) -> np.ndarray:
"""
Shift pitch without changing duration using phase vocoder
pitch_factor > 1.0 = higher pitch, < 1.0 = lower pitch
"""
if pitch_factor == 1.0:
return audio
try:
import librosa
# Pitch shift in semitones
semitones = 12 * np.log2(pitch_factor)
shifted = librosa.effects.pitch_shift(
audio.astype(np.float32), sr=sample_rate, n_steps=semitones
)
return shifted
except ImportError:
# Fallback: simple resampling-based pitch shift (changes duration slightly)
from scipy import signal
# Resample to change pitch, then resample back to original length
stretched = signal.resample(audio, int(len(audio) / pitch_factor))
return signal.resample(stretched, len(audio))
@staticmethod
def apply_speed_change(
audio: np.ndarray, sample_rate: int, speed_factor: float
) -> np.ndarray:
"""
Change speed/tempo without changing pitch
speed_factor > 1.0 = faster, < 1.0 = slower
"""
if speed_factor == 1.0:
return audio
try:
import librosa
# Time stretch
stretched = librosa.effects.time_stretch(
audio.astype(np.float32), rate=speed_factor
)
return stretched
except ImportError:
# Fallback: simple resampling (will also change pitch)
from scipy import signal
target_length = int(len(audio) / speed_factor)
return signal.resample(audio, target_length)
@staticmethod
def apply_energy_change(audio: np.ndarray, energy_factor: float) -> np.ndarray:
"""
Modify audio energy/volume
energy_factor > 1.0 = louder, < 1.0 = softer
"""
if energy_factor == 1.0:
return audio
# Apply gain with soft clipping to avoid distortion
modified = audio * energy_factor
# Soft clip using tanh for natural sound
if energy_factor > 1.0:
max_val = np.max(np.abs(modified))
if max_val > 0.95:
modified = np.tanh(modified * 2) * 0.95
return modified
@staticmethod
def apply_style(
audio: np.ndarray,
sample_rate: int,
speed: float = 1.0,
pitch: float = 1.0,
energy: float = 1.0,
) -> np.ndarray:
"""Apply all style modifications"""
result = audio
# Apply in order: pitch -> speed -> energy
if pitch != 1.0:
result = StyleProcessor.apply_pitch_shift(result, sample_rate, pitch)
if speed != 1.0:
result = StyleProcessor.apply_speed_change(result, sample_rate, speed)
if energy != 1.0:
result = StyleProcessor.apply_energy_change(result, energy)
return result
@staticmethod
def get_preset(preset_name: str) -> Dict[str, float]:
"""Get style parameters from preset name"""
return STYLE_PRESETS.get(preset_name, STYLE_PRESETS["default"])
class TTSEngine:
"""
Multi-lingual TTS Engine using SYSPIN VITS models
Supports 11 Indian languages with male/female voices:
- Hindi, Bengali, Marathi, Telugu, Kannada
- Bhojpuri, Chhattisgarhi, Maithili, Magahi, English
- Gujarati (via Facebook MMS)
Features:
- Style/prosody control (pitch, speed, energy)
- Preset styles (happy, sad, calm, excited, etc.)
- JIT traced models (.pt) and Coqui TTS checkpoints (.pth)
"""
def __init__(
self,
models_dir: str = MODELS_DIR,
device: str = "auto",
preload_voices: Optional[List[str]] = None,
):
"""
Initialize TTS Engine
Args:
models_dir: Directory containing downloaded models
device: Device to run inference on ('cpu', 'cuda', 'mps', or 'auto')
preload_voices: List of voice keys to preload into memory
"""
self.models_dir = Path(models_dir)
self.device = self._get_device(device)
# Model cache - JIT traced models (.pt)
self._models: Dict[str, torch.jit.ScriptModule] = {}
self._tokenizers: Dict[str, TTSTokenizer] = {}
# Coqui TTS models cache (.pth checkpoints)
self._coqui_models: Dict[str, Any] = {} # Stores Synthesizer objects
# MMS models cache (separate handling)
self._mms_models: Dict[str, Any] = {}
self._mms_tokenizers: Dict[str, Any] = {}
# Downloader
self.downloader = ModelDownloader(models_dir)
# Text normalizer
self.normalizer = TextNormalizer()
# Style processor
self.style_processor = StyleProcessor()
# Preload specified voices
if preload_voices:
for voice in preload_voices:
self.load_voice(voice)
logger.info(f"TTS Engine initialized on device: {self.device}")
def _get_device(self, device: str) -> torch.device:
"""Determine the best device for inference"""
if device == "auto":
if torch.cuda.is_available():
return torch.device("cuda")
# MPS has compatibility issues with some TorchScript models
# Using CPU for now - still fast on Apple Silicon
# elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
# return torch.device("mps")
else:
return torch.device("cpu")
return torch.device(device)
def load_voice(self, voice_key: str, download_if_missing: bool = True) -> bool:
"""
Load a voice model into memory
Args:
voice_key: Key from LANGUAGE_CONFIGS (e.g., 'hi_male')
download_if_missing: Download model if not found locally
Returns:
True if loaded successfully
"""
# Check if already loaded
if voice_key in self._models or voice_key in self._coqui_models:
return True
if voice_key not in LANGUAGE_CONFIGS:
raise ValueError(f"Unknown voice: {voice_key}")
config = LANGUAGE_CONFIGS[voice_key]
model_dir = self.models_dir / voice_key
# Check if model exists, download if needed
if not model_dir.exists():
if download_if_missing:
logger.info(f"Model not found, downloading {voice_key}...")
self.downloader.download_model(voice_key)
else:
raise FileNotFoundError(f"Model directory not found: {model_dir}")
# Check for Coqui TTS checkpoint (.pth) vs JIT traced model (.pt)
pth_files = list(model_dir.glob("*.pth"))
pt_files = list(model_dir.glob("*.pt"))
if pth_files:
# Load as Coqui TTS checkpoint
return self._load_coqui_voice(voice_key, model_dir, pth_files[0])
elif pt_files:
# Load as JIT traced model
return self._load_jit_voice(voice_key, model_dir, pt_files[0])
else:
raise FileNotFoundError(f"No .pt or .pth model file found in {model_dir}")
def _load_jit_voice(
self, voice_key: str, model_dir: Path, model_path: Path
) -> bool:
"""
Load a JIT traced VITS model (.pt file)
"""
# Load tokenizer
chars_path = model_dir / "chars.txt"
if chars_path.exists():
tokenizer = TTSTokenizer.from_chars_file(str(chars_path))
else:
# Try to find chars file
chars_files = list(model_dir.glob("*chars*.txt"))
if chars_files:
tokenizer = TTSTokenizer.from_chars_file(str(chars_files[0]))
else:
raise FileNotFoundError(f"No chars.txt found in {model_dir}")
# Load model
logger.info(f"Loading JIT model from {model_path}")
model = torch.jit.load(str(model_path), map_location=self.device)
model.eval()
# Cache model and tokenizer
self._models[voice_key] = model
self._tokenizers[voice_key] = tokenizer
logger.info(f"Loaded JIT voice: {voice_key}")
return True
def _load_coqui_voice(
self, voice_key: str, model_dir: Path, checkpoint_path: Path
) -> bool:
"""
Load a Coqui TTS checkpoint model (.pth file)
"""
config_path = model_dir / "config.json"
if not config_path.exists():
raise FileNotFoundError(f"No config.json found in {model_dir}")
try:
from TTS.utils.synthesizer import Synthesizer
logger.info(f"Loading Coqui TTS checkpoint from {checkpoint_path}")
# Create synthesizer with checkpoint and config
use_cuda = self.device.type == "cuda"
synthesizer = Synthesizer(
tts_checkpoint=str(checkpoint_path),
tts_config_path=str(config_path),
use_cuda=use_cuda,
)
# Cache synthesizer
self._coqui_models[voice_key] = synthesizer
logger.info(f"Loaded Coqui voice: {voice_key}")
return True
except ImportError:
raise ImportError(
"Coqui TTS library not installed. " "Install it with: pip install TTS"
)
def _synthesize_coqui(self, text: str, voice_key: str) -> Tuple[np.ndarray, int]:
"""
Synthesize using Coqui TTS model (for Bhojpuri etc.)
"""
if voice_key not in self._coqui_models:
self.load_voice(voice_key)
synthesizer = self._coqui_models[voice_key]
config = LANGUAGE_CONFIGS[voice_key]
# Generate audio
wav = synthesizer.tts(text)
# Convert to numpy array
audio_np = np.array(wav, dtype=np.float32)
sample_rate = synthesizer.output_sample_rate
return audio_np, sample_rate
def _load_mms_voice(self, voice_key: str) -> bool:
"""
Load Facebook MMS model for Gujarati
"""
if voice_key in self._mms_models:
return True
config = LANGUAGE_CONFIGS[voice_key]
logger.info(f"Loading MMS model: {config.hf_model_id}")
try:
from transformers import VitsModel, AutoTokenizer
# Load model and tokenizer from HuggingFace
model = VitsModel.from_pretrained(config.hf_model_id)
tokenizer = AutoTokenizer.from_pretrained(config.hf_model_id)
model = model.to(self.device)
model.eval()
self._mms_models[voice_key] = model
self._mms_tokenizers[voice_key] = tokenizer
logger.info(f"Loaded MMS voice: {voice_key}")
return True
except Exception as e:
logger.error(f"Failed to load MMS model: {e}")
raise
def _synthesize_mms(self, text: str, voice_key: str) -> Tuple[np.ndarray, int]:
"""
Synthesize using Facebook MMS model (for Gujarati)
"""
if voice_key not in self._mms_models:
self._load_mms_voice(voice_key)
model = self._mms_models[voice_key]
tokenizer = self._mms_tokenizers[voice_key]
config = LANGUAGE_CONFIGS[voice_key]
# Tokenize
inputs = tokenizer(text, return_tensors="pt")
inputs = {k: v.to(self.device) for k, v in inputs.items()}
# Generate
with torch.no_grad():
output = model(**inputs)
# Get audio
audio = output.waveform.squeeze().cpu().numpy()
return audio, config.sample_rate
def unload_voice(self, voice_key: str):
"""Unload a voice to free memory"""
if voice_key in self._models:
del self._models[voice_key]
del self._tokenizers[voice_key]
if voice_key in self._coqui_models:
del self._coqui_models[voice_key]
if voice_key in self._mms_models:
del self._mms_models[voice_key]
del self._mms_tokenizers[voice_key]
torch.cuda.empty_cache() if self.device.type == "cuda" else None
logger.info(f"Unloaded voice: {voice_key}")
def synthesize(
self,
text: str,
voice: str = "hi_male",
speed: float = 1.0,
pitch: float = 1.0,
energy: float = 1.0,
style: Optional[str] = None,
normalize_text: bool = True,
) -> TTSOutput:
"""
Synthesize speech from text with style control
Args:
text: Input text to synthesize
voice: Voice key (e.g., 'hi_male', 'bn_female', 'gu_mms')
speed: Speech speed multiplier (0.5-2.0)
pitch: Pitch multiplier (0.5-2.0), >1 = higher
energy: Energy/volume multiplier (0.5-2.0)
style: Style preset name (e.g., 'happy', 'sad', 'calm')
normalize_text: Whether to apply text normalization
Returns:
TTSOutput with audio array and metadata
"""
# Apply style preset if specified
if style and style in STYLE_PRESETS:
preset = STYLE_PRESETS[style]
speed = speed * preset["speed"]
pitch = pitch * preset["pitch"]
energy = energy * preset["energy"]
config = LANGUAGE_CONFIGS[voice]
# Normalize text
if normalize_text:
text = self.normalizer.clean_text(text, config.code)
# Check if this is an MMS model (Gujarati)
if "mms" in voice:
audio_np, sample_rate = self._synthesize_mms(text, voice)
# Check if this is a Coqui TTS model (Bhojpuri etc.)
elif voice in self._coqui_models:
audio_np, sample_rate = self._synthesize_coqui(text, voice)
else:
# Try to load the voice (will determine JIT vs Coqui)
if voice not in self._models and voice not in self._coqui_models:
self.load_voice(voice)
# Check again after loading
if voice in self._coqui_models:
audio_np, sample_rate = self._synthesize_coqui(text, voice)
else:
# Use JIT model (SYSPIN models)
model = self._models[voice]
tokenizer = self._tokenizers[voice]
# Tokenize
token_ids = tokenizer.text_to_ids(text)
x = torch.from_numpy(np.array(token_ids)).unsqueeze(0).to(self.device)
# Generate audio
with torch.no_grad():
audio = model(x)
audio_np = audio.squeeze().cpu().numpy()
sample_rate = config.sample_rate
# Apply style modifications (pitch, speed, energy)
audio_np = self.style_processor.apply_style(
audio_np, sample_rate, speed=speed, pitch=pitch, energy=energy
)
# Calculate duration
duration = len(audio_np) / sample_rate
return TTSOutput(
audio=audio_np,
sample_rate=sample_rate,
duration=duration,
voice=voice,
text=text,
style=style,
)
def synthesize_to_file(
self,
text: str,
output_path: str,
voice: str = "hi_male",
speed: float = 1.0,
pitch: float = 1.0,
energy: float = 1.0,
style: Optional[str] = None,
normalize_text: bool = True,
) -> str:
"""
Synthesize speech and save to file
Args:
text: Input text to synthesize
output_path: Path to save audio file
voice: Voice key
speed: Speech speed multiplier
pitch: Pitch multiplier
energy: Energy multiplier
style: Style preset name
normalize_text: Whether to apply text normalization
Returns:
Path to saved file
"""
import soundfile as sf
output = self.synthesize(
text, voice, speed, pitch, energy, style, normalize_text
)
sf.write(output_path, output.audio, output.sample_rate)
logger.info(f"Saved audio to {output_path} (duration: {output.duration:.2f}s)")
return output_path
def get_loaded_voices(self) -> List[str]:
"""Get list of currently loaded voices"""
return (
list(self._models.keys())
+ list(self._coqui_models.keys())
+ list(self._mms_models.keys())
)
def get_available_voices(self) -> Dict[str, Dict]:
"""Get all available voices with their status"""
voices = {}
for key, config in LANGUAGE_CONFIGS.items():
is_mms = "mms" in key
model_dir = self.models_dir / key
# Determine model type
if is_mms:
model_type = "mms"
elif model_dir.exists() and list(model_dir.glob("*.pth")):
model_type = "coqui"
else:
model_type = "vits"
voices[key] = {
"name": config.name,
"code": config.code,
"gender": (
"male"
if "male" in key
else ("female" if "female" in key else "neutral")
),
"loaded": key in self._models
or key in self._coqui_models
or key in self._mms_models,
"downloaded": is_mms or self.downloader.get_model_path(key) is not None,
"type": model_type,
}
return voices
def get_style_presets(self) -> Dict[str, Dict]:
"""Get available style presets"""
return STYLE_PRESETS
def batch_synthesize(
self, texts: List[str], voice: str = "hi_male", speed: float = 1.0
) -> List[TTSOutput]:
"""Synthesize multiple texts"""
return [self.synthesize(text, voice, speed) for text in texts]
# Convenience function
def synthesize(
text: str, voice: str = "hi_male", output_path: Optional[str] = None
) -> Union[TTSOutput, str]:
"""
Quick synthesis function
Args:
text: Text to synthesize
voice: Voice key
output_path: If provided, saves to file and returns path
Returns:
TTSOutput if no output_path, else path to saved file
"""
engine = TTSEngine()
if output_path:
return engine.synthesize_to_file(text, output_path, voice)
return engine.synthesize(text, voice)