Piper TTS

Converts input text to natural-sounding speech using VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech) neural networks exported to ONNX format for CPU-efficient inference. The C++ core engine loads pre-trained ONNX models and executes the full synthesis pipeline (text→phonemes→mel-spectrogram→waveform) locally without cloud dependencies, optimized for edge devices like Raspberry Pi 4 with minimal memory footprint and latency.

Processes raw text input through language-specific normalization rules and converts graphemes to phoneme sequences using espeak-ng backend, handling abbreviations, numbers, punctuation, and language-specific phonetic rules. The pipeline supports 30+ languages with language-specific phoneme inventories defined in voice configuration JSON files, enabling accurate phonetic representation for downstream neural synthesis.

Provides Docker configuration and build scripts for containerizing Piper as a self-contained service, enabling reproducible deployment across different environments. The container includes the C++ engine, Python API, HTTP server, and voice models, with environment variable configuration for voice selection and server parameters.

Includes benchmarking tools and optimization techniques for measuring and improving inference performance on resource-constrained devices, including model quantization, batch processing analysis, and latency profiling. The system profiles synthesis time, memory usage, and CPU utilization across different device types (Raspberry Pi, Jetson, etc.) to guide model selection and optimization.

Loads VITS models trained on multiple speakers and selects speaker embeddings at inference time based on voice configuration mappings, enabling a single model to synthesize speech with different voice characteristics (pitch, timbre, speaking style). The speaker selection is controlled via speaker ID or speaker name lookup in the voice configuration JSON, allowing dynamic voice switching without model reloading.

Synthesizes speech as continuous PCM audio streams with configurable output sample rates (22050Hz, 44100Hz, 48000Hz) and bit depths (float32, int16), supporting real-time audio playback and file writing. The synthesis engine generates mel-spectrograms from phoneme sequences and converts them to waveform samples via neural vocoder, with streaming output enabling low-latency playback on resource-constrained devices without buffering entire audio in memory.

Provides a CLI tool that accepts text input (from stdin or file arguments) and synthesizes speech to WAV files, supporting voice selection, speaker selection for multi-speaker models, and output file specification. The CLI wraps the C++ core engine and handles file I/O, argument parsing, and error handling, making Piper accessible without programming knowledge.

Exposes Piper's TTS engine through a Python module with classes for voice loading, synthesis, and audio output, enabling integration into Python applications. The API manages ONNX model lifecycle (loading, caching), handles phonemization and synthesis in Python, and provides generator-based streaming for memory-efficient processing of large text batches.

Runs Piper as a network service exposing REST endpoints for text-to-speech synthesis, enabling remote clients to request speech synthesis over HTTP. The server manages model loading, request queuing, and concurrent synthesis requests, supporting voice and speaker selection via query parameters or JSON request bodies, with audio returned as WAV or raw PCM.

Provides utilities to discover, download, and manage voice models from the Hugging Face model hub, with automatic caching and version management. The system maintains a local voice directory with downloaded .onnx model files and .onnx.json configuration files, supporting model listing, updates, and cleanup without manual file management.

Provides end-to-end training infrastructure for creating custom voice models from audio recordings and transcripts, including data preparation, model training with VITS architecture, and ONNX export. The pipeline handles audio preprocessing, phoneme alignment, speaker embedding training, and model optimization for edge device inference, enabling users to train domain-specific or custom voices.

Manages voice-specific metadata in JSON configuration files (.onnx.json) including phoneme inventory, speaker ID mappings, synthesis parameters (noise scale, length scale), and model architecture details. The configuration system enables flexible voice customization without model retraining, supporting per-voice phoneme sets, speaker naming, and synthesis quality tuning.