feat: Add Piper TTS integration with pitch control support
- Created PiperTTSService with pitch shifting (-12 to +12 semitones) - Created UnifiedTTSService to manage multiple TTS engines - Added engine selection endpoints (GET/POST /api/tts/engines) - Added voices listing endpoint (GET /api/tts/voices) - Added pitch_adjustments parameter to TTS generation - Created install_piper.sh for easy Piper installation - Updated backend README with multi-engine documentation - Created comprehensive PIPER_TTS_GUIDE.md - Added TTS_ENGINE environment variable for engine selection - Updated requirements.txt with piper-tts and pydub This allows users to choose between Kokoro (high quality) and Piper (fast with pitch control) TTS engines at runtime.
This commit is contained in:
219
backend/services/piper_tts_service.py
Normal file
219
backend/services/piper_tts_service.py
Normal file
@@ -0,0 +1,219 @@
|
||||
"""
|
||||
Piper TTS service for generating multi-voice audiobooks.
|
||||
"""
|
||||
import logging
|
||||
import io
|
||||
import os
|
||||
import subprocess
|
||||
from typing import Dict, List, Tuple, Optional
|
||||
import numpy as np
|
||||
import soundfile as sf
|
||||
from pydub import AudioSegment
|
||||
from pydub.effects import speedup
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class PiperTTSService:
|
||||
"""
|
||||
Service for text-to-speech synthesis using Piper TTS.
|
||||
|
||||
Piper is a fast, local neural TTS system that supports multiple voices
|
||||
and allows pitch modification through audio processing.
|
||||
"""
|
||||
|
||||
# Default voice for unmapped roles
|
||||
DEFAULT_VOICE = "en_US-lessac-medium"
|
||||
|
||||
# Sample rate for Piper TTS (typically 22050)
|
||||
SAMPLE_RATE = 22050
|
||||
|
||||
# Available Piper voices with their characteristics
|
||||
AVAILABLE_VOICES = {
|
||||
# English voices
|
||||
"en_US-lessac-medium": {"lang": "en", "gender": "male", "quality": "medium"},
|
||||
"en_US-libritts-high": {"lang": "en", "gender": "neutral", "quality": "high"},
|
||||
"en_US-amy-medium": {"lang": "en", "gender": "female", "quality": "medium"},
|
||||
"en_GB-alan-medium": {"lang": "en", "gender": "male", "quality": "medium"},
|
||||
"en_GB-alba-medium": {"lang": "en", "gender": "female", "quality": "medium"},
|
||||
}
|
||||
|
||||
def __init__(self):
|
||||
"""Initializes Piper TTS service."""
|
||||
self.models_dir = os.path.join(os.path.dirname(__file__), "..", "piper_models")
|
||||
os.makedirs(self.models_dir, exist_ok=True)
|
||||
logger.info("PiperTTSService initialized")
|
||||
|
||||
def _check_piper_installed(self) -> bool:
|
||||
"""Check if Piper is installed and accessible."""
|
||||
try:
|
||||
result = subprocess.run(
|
||||
["piper", "--version"],
|
||||
capture_output=True,
|
||||
text=True,
|
||||
timeout=5
|
||||
)
|
||||
return result.returncode == 0
|
||||
except (subprocess.SubprocessError, FileNotFoundError):
|
||||
return False
|
||||
|
||||
def _download_voice_model(self, voice: str) -> bool:
|
||||
"""
|
||||
Download a Piper voice model if not already present.
|
||||
|
||||
Args:
|
||||
voice: Voice ID to download
|
||||
|
||||
Returns:
|
||||
True if model is available, False otherwise
|
||||
"""
|
||||
model_path = os.path.join(self.models_dir, f"{voice}.onnx")
|
||||
config_path = os.path.join(self.models_dir, f"{voice}.onnx.json")
|
||||
|
||||
# Check if model already exists
|
||||
if os.path.exists(model_path) and os.path.exists(config_path):
|
||||
return True
|
||||
|
||||
logger.info(f"Downloading Piper voice model: {voice}")
|
||||
|
||||
# Download using piper's built-in download
|
||||
try:
|
||||
subprocess.run(
|
||||
["piper", "--model", voice, "--download-dir", self.models_dir],
|
||||
check=True,
|
||||
timeout=60
|
||||
)
|
||||
return os.path.exists(model_path)
|
||||
except subprocess.SubprocessError as e:
|
||||
logger.error(f"Failed to download voice model {voice}: {e}")
|
||||
return False
|
||||
|
||||
def synthesize_with_pitch(
|
||||
self,
|
||||
text: str,
|
||||
voice: str,
|
||||
speed: float = 1.0,
|
||||
pitch_shift: int = 0
|
||||
) -> np.ndarray:
|
||||
"""
|
||||
Synthesizes text with Piper and applies pitch shifting.
|
||||
|
||||
Args:
|
||||
text: Text to synthesize
|
||||
voice: Piper voice ID
|
||||
speed: Playback speed (0.5-2.0)
|
||||
pitch_shift: Semitones to shift pitch (-12 to +12)
|
||||
|
||||
Returns:
|
||||
Audio samples as numpy array
|
||||
"""
|
||||
# Ensure voice model is available
|
||||
if not self._download_voice_model(voice):
|
||||
logger.warning(f"Voice {voice} not available, using default")
|
||||
voice = self.DEFAULT_VOICE
|
||||
self._download_voice_model(voice)
|
||||
|
||||
model_path = os.path.join(self.models_dir, f"{voice}.onnx")
|
||||
|
||||
# Generate audio with Piper
|
||||
try:
|
||||
# Run piper command
|
||||
process = subprocess.Popen(
|
||||
["piper", "--model", model_path, "--output-raw"],
|
||||
stdin=subprocess.PIPE,
|
||||
stdout=subprocess.PIPE,
|
||||
stderr=subprocess.PIPE
|
||||
)
|
||||
|
||||
# Send text and get audio
|
||||
stdout, stderr = process.communicate(input=text.encode('utf-8'), timeout=30)
|
||||
|
||||
if process.returncode != 0:
|
||||
logger.error(f"Piper synthesis failed: {stderr.decode()}")
|
||||
return np.array([], dtype=np.float32)
|
||||
|
||||
# Convert raw audio to numpy array
|
||||
audio_data = np.frombuffer(stdout, dtype=np.int16)
|
||||
audio_float = audio_data.astype(np.float32) / 32768.0
|
||||
|
||||
# Apply pitch shift and speed change using pydub
|
||||
if pitch_shift != 0 or speed != 1.0:
|
||||
audio_float = self._apply_audio_effects(
|
||||
audio_float,
|
||||
self.SAMPLE_RATE,
|
||||
speed,
|
||||
pitch_shift
|
||||
)
|
||||
|
||||
return audio_float
|
||||
|
||||
except subprocess.TimeoutExpired:
|
||||
logger.error("Piper synthesis timed out")
|
||||
return np.array([], dtype=np.float32)
|
||||
except Exception as e:
|
||||
logger.error(f"Error in Piper synthesis: {e}")
|
||||
return np.array([], dtype=np.float32)
|
||||
|
||||
def _apply_audio_effects(
|
||||
self,
|
||||
audio: np.ndarray,
|
||||
sample_rate: int,
|
||||
speed: float,
|
||||
pitch_shift: int
|
||||
) -> np.ndarray:
|
||||
"""
|
||||
Apply speed and pitch effects to audio using pydub.
|
||||
|
||||
Args:
|
||||
audio: Audio samples
|
||||
sample_rate: Sample rate
|
||||
speed: Speed multiplier
|
||||
pitch_shift: Semitones to shift
|
||||
|
||||
Returns:
|
||||
Processed audio samples
|
||||
"""
|
||||
try:
|
||||
# Convert numpy to pydub AudioSegment
|
||||
audio_int16 = (audio * 32768.0).astype(np.int16)
|
||||
audio_bytes = audio_int16.tobytes()
|
||||
|
||||
audio_segment = AudioSegment(
|
||||
audio_bytes,
|
||||
frame_rate=sample_rate,
|
||||
sample_width=2,
|
||||
channels=1
|
||||
)
|
||||
|
||||
# Apply pitch shift
|
||||
if pitch_shift != 0:
|
||||
# Pitch shift by changing frame rate then resampling
|
||||
new_sample_rate = int(sample_rate * (2.0 ** (pitch_shift / 12.0)))
|
||||
audio_segment = audio_segment._spawn(
|
||||
audio_segment.raw_data,
|
||||
overrides={'frame_rate': new_sample_rate}
|
||||
)
|
||||
audio_segment = audio_segment.set_frame_rate(sample_rate)
|
||||
|
||||
# Apply speed change
|
||||
if speed != 1.0:
|
||||
audio_segment = speedup(audio_segment, playback_speed=speed)
|
||||
|
||||
# Convert back to numpy
|
||||
samples = np.array(audio_segment.get_array_of_samples())
|
||||
audio_float = samples.astype(np.float32) / 32768.0
|
||||
|
||||
return audio_float
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Error applying audio effects: {e}")
|
||||
return audio
|
||||
|
||||
def get_available_voices(self) -> Dict[str, Dict[str, str]]:
|
||||
"""
|
||||
Returns dictionary of available Piper voices.
|
||||
|
||||
Returns:
|
||||
Dict mapping voice IDs to their characteristics
|
||||
"""
|
||||
return self.AVAILABLE_VOICES.copy()
|
||||
Reference in New Issue
Block a user