FastTTS 0.1.6 [ALPHA-2026-10-17] — Unified, Zero-Bloat TTS Backend Orchestration for Java

⚡ Connect multiple TTS backends with a unified interface — Minimalist text-to-speech orchestration supporting offline neural engines, system voices, zero-shot voice cloning, and cloud providers.

FastTTS is a lightweight, framework-agnostic TTS engine designed to provide unified access to multiple text-to-speech backends with zero framework bloat. It supports Kokoro Studio (Sherpa-ONNX), OmniVoice (Diffusion Zero-Shot Cloning), Pocket TTS (Sherpa-ONNX INT8), Piper (offline ONNX), Dia2 (GGUF Dialogue), Windows SAPI (system), Hugging Face Spaces, ElevenLabs, and Deepgram through a single clean API.

Watch Demo (YouTube) | Watch JMH Benchmark (YouTube)

Quick Start — Example

import fasttts.FastTTS;
import fasttts.backends.kokoro.KokoroBackend;
import fasttts.backends.kokoro.KokoroVoice;
import fasttts.backends.pockettts.PocketTTSBackend;
import fasttts.backends.piper.PiperBackend;
import fasttts.backends.windows.WindowsTTSBackend;
import fasttts.core.FastTTSAudio;

public class Demo {
    public static void main(String[] args) throws Exception {
        // FastTTS with pre-registered auto-discovered engines
        try (FastTTS tts = FastTTS.standard()) {
            
            // 1. Windows Native (ultra-fast 150ms fallback, 0.03x RTF)
            FastTTSAudio winAudio = tts.speak("windows", "Windows SAPI native speech is online.", null, null);

            // 2. Kokoro Studio (studio-grade neural voice)
            FastTTSAudio kokoroAudio = tts.speak("kokoro", "Kokoro studio neural speech is active.", KokoroVoice.AF_BELLA, null);
            
            // 3. Matcha-TTS (Flow-Matching ultra-fast neural speech, 0.09x RTF)
            FastTTSAudio matchaAudio = tts.speak("matcha", "Matcha flow matching is running at extreme speed.", null, null);

            // 4. VITS (Deterministic low-latency neural TTS, 0.09x RTF)
            FastTTSAudio vitsAudio = tts.speak("vits", "VITS deterministic synthesis is online.", null, null);

            // 5. Piper (offline German/English neural speech)
            FastTTSAudio piperAudio = tts.speak("piper", "Piper offline neural speech is active.", null, null);
            
            // 6. Pocket TTS (ultra-compact INT8 CPU neural engine)
            FastTTSAudio pocketAudio = tts.speak("pocket", "Pocket TTS compact neural voice is ready.", null, null);
        }
    }
}

Why FastTTS?

Traditional TTS libraries force developers into heavyweight Python dependencies, complex cloud API integrations, or platform-specific code. FastTTS provides:

100% Native JVM Pipeline — Orchestrates multiple TTS backends in a single JVM process with unified interface.
Offline Neural & Cloud Support — Seamlessly switch between local CPU neural models (Pocket TTS, Piper), system voices (Windows SAPI), and cloud providers (ElevenLabs, Deepgram).
Model Agnostic — Works with quantized INT8 ONNX models, system voices, and cloud APIs through the same FastTTSBackend interface.
Zero Configuration Overlap — Integrates seamlessly with existing FastJava ecosystem libraries.

FastTTS unifies local neural inference, system voices, and cloud TTS into a single framework-agnostic API:

Feature	MaryTTS (Legacy Java)	Python Coqui/Piper Subprocess	FastTTS
Synthesis Backends	Obsolete Java voices only	Single CLI wrapper	Unified (Kokoro, Matcha-TTS, VITS, OmniVoice, Pocket TTS, Piper, F5-TTS, Dia2, SAPI, Cloud)
Offline Privacy	Yes (Robotic sound)	Yes (Heavy Python runtime)	100% Offline (Kokoro, Matcha-TTS, VITS, Pocket TTS, Piper, F5-TTS, Dia2, Windows SAPI)
Startup / Synthesis	800–2500 ms (Heap heavy)	2000–5000 ms (Process spawn)	~138 ms (Windows SAPI) / ~378 ms (VITS) / ~570 ms (Matcha) / ~696 ms (Piper)
Voice Quality	Synthetic 2000s robotic	Neural quality	Studio-Grade Neural (Kokoro, Matcha Flow-Matching, OmniVoice Zero-Shot)
Barge-In Ready	Difficult to interrupt	Stalled CLI process	Instant Cancellation via FastVAD & WASAPI
Dependencies	Massive legacy JARs	Python 3 + Pip dependencies	Pure Java 17+ backed by FastCore

Key Features

🎭 Multiple Backend Support — Unified interface for Kokoro Studio (ONNX), Matcha-TTS (Flow Matching), VITS (Deterministic Neural), OmniVoice (Zero-Shot Diffusion), Pocket TTS (Sherpa-ONNX INT8), Piper (offline ONNX), F5-TTS (Flow-Matching Voice Clone), Dia2 (GGUF Dialogue), Windows SAPI (system), and Cloud APIs.
📱 100% Offline Capable — High-fidelity local speech synthesis without internet connections or cloud subscription costs.
🌊 Flow Matching & Non-Autoregressive Synthesis — State-of-the-art optimal-transport Flow Matching via Matcha-TTS running at >10x real-time speed on CPU.
🎙️ Zero-Shot Voice Cloning — Clone custom voices on the fly from reference WAV audio with OmniVoice and F5-TTS.
💬 Multi-Speaker Dialogue — Native support for scripted dialogue ([S1], [S2]) via Dia2 GGUF models.
⚡ Real-Time Factor (RTF) Acceleration — Hardware SIMD acceleration with 0.03x RTF on SAPI, 0.09x RTF on VITS and Matcha-TTS, and 0.25x RTF on Piper.
🧵 Complete Lifecycle Control — AutoCloseable orchestration with pre-warming (prepare()) and deterministic process teardown.

Real-World Use Cases

🗣️ Conversational Voice AI Assistants: Low-latency neural speech synthesis for AI chatbots and desktop voice assistants.
📖 Audiobook & Content Reader: High-speed offline speech synthesis for document reading and accessibility tools.
📢 In-App & Game Audio Notifications: Real-time voice announcements with zero Garbage Collection latency impact.
🌐 Multi-Language Accessibility Engines: Seamlessly switch between local Piper voices, Kokoro studio voices, and cloud APIs.

Architecture Overview

       ┌───────────────────────────────────────────────────────────────┐
       │             FastTTS (Public Orchestration Facade)             │
       │     AutoCloseable  •  Streaming Callback  •  Voice Router     │
       └───────────────────────────────┬───────────────────────────────┘
                                       │
        ┌──────────────────────────────┼──────────────────────────────┐
        ▼                              ▼                              ▼
 ┌──────────────┐              ┌──────────────┐              ┌────────────────┐
 │ System SAPI  │              │ Offline ONNX │              │  Cloud Engines │
 │ fasttts.dll  │              │ Sherpa/Piper │              │ ElevenLabs etc │
 └──────┬───────┘              └──────┬───────┘              └────────┬───────┘
        │                             │                               │
        │ (~152 ms, 0.03x RTF)        │ (AVX2 INT8 / Zero-Shot)       │ (Remote REST)
        ▼                             ▼                               ▼
 ┌────────────────────────────────────────────────────────────────────────────┐
 │                  FastTTSAudio (Standard Linear PCM Buffer)                 │
 │            Sample Rate  •  Channels (Mono/Stereo)  •  Raw Bytes            │
 └─────────────────────────────────────┬──────────────────────────────────────┘
                                       │
                                       ▼
 ┌────────────────────────────────────────────────────────────────────────────┐
 │         FastJava AVX2 Audio DSP & Low-Latency Playback Substrate           │
 │       FastAudioProcess (SIMD DSP)  •  FastAudioPlayer (Direct WASAPI)      │
 └────────────────────────────────────────────────────────────────────────────┘

Core Backend Engines

🔊 Windows SAPI Backend: System-native text-to-speech using Windows Speech API via AVX2-optimized asttts.dll (~152 ms latency, 0.03x RTF).
🍵 Matcha-TTS (Flow-Matching): High-speed Flow Matching neural speech via sherpa-onnx and Vocos vocoder at 22.05 kHz (~0.09x RTF, > 10× real-time).
⚡ VITS (Piper ONNX): Deterministic low-latency variational inference neural speech (~0.09x RTF).
🎙️ Kokoro Studio (ONNX): Studio-grade offline neural TTS powered by Sherpa-ONNX (kokoro-en-v0_19) with high-fidelity voices (�f_bella, �m_adam, �f_emma, �m_george).
⚡ Piper Neural (ONNX): Ultra-fast offline neural TTS using the Piper ONNX engine (piper.exe) with multi-directory auto-discovery (~848 ms latency, 0.25x RTF).
📱 Pocket TTS (ONNX): Ultra-compact, low-latency offline CPU neural text-to-speech engine powered by Sherpa-ONNX with quantized INT8 models (0.71x - 0.96x RTF).
🧬 F5-TTS (Flow-Matching): Advanced non-autoregressive flow-matching zero-shot voice cloning via �udiocpp_cli.exe.
🎭 OmniVoice (Diffusion): Multilingual zero-shot voice cloning diffusion model supporting 600+ languages from short reference audio.
💬 Dia2 Dialogue (GGUF): Quantized GGUF dialogue model (dia2-2b-q4_k.gguf) with multi-speaker turn parsing and memory page cache warming.
🌐 Cloud Backends: Cloud APIs (Hugging Face Spaces, ElevenLabs, Deepgram, OpenAI) with hardened JSON escaping and automatic token resolution.

Performance Benchmarks

Directly benchmarked across all engines via un-demo.bat benchmark with audio playback through FastAudioPlayer (WASAPI):

Engine / Backend	Voice / Model	Lang	Synth Time	Audio Length	Real-Time Factor (RTF)	Buffer Size	Status
Windows SAPI	OneCore/Default	en-US	152 ms	5,044 ms	0.030x	434.5 KB	✅ Active
Matcha-TTS (FlowMatch)	matcha-icefall	en-US	320 ms	3,550 ms	0.090x	156.4 KB	✅ Active
VITS Neural (ONNX)	�my-low	en-US	315 ms	3,500 ms	0.090x	154.2 KB	✅ Active
Piper Neural (ONNX)	horsten-medium	de-DE	848 ms	3,405 ms	0.249x	146.7 KB	✅ Active
Pocket TTS (ONNX)	int8-bria	en-US	3,473 ms	3,600 ms	0.965x	168.8 KB	✅ Active
Kokoro Studio (ONNX)	�m_george	en-GB	3,935 ms	4,373 ms	0.900x	205.0 KB	✅ Active
Kokoro Studio (ONNX)	�f_emma	en-GB	4,281 ms	4,843 ms	0.884x	227.0 KB	✅ Active
Kokoro Studio (ONNX)	�m_adam	en-US	4,053 ms	4,187 ms	0.968x	196.3 KB	✅ Active
Kokoro Studio (ONNX)	�f_bella	en-US	4,479 ms	4,363 ms	1.027x	204.6 KB	✅ Active
OmniVoice (Diffusion)	�ria (Zero-Shot)	en-US	54,413 ms	4,040 ms	13.469x	189.4 KB	✅ Active
F5-TTS (FlowMatch)	5-v1-base	en-US	—	—	—	—	📦 Model Ready (999 MB)
Dia2 Dialogue (GGUF)	dia2-2b-q4_k	en-US	—	—	—	—	📦 Model Ready (1.2 GB)
Hugging Face API	mms-tts-eng	en-US	—	—	—	—	🔑 Cloud (Needs Token)

Real-Time Factor (RTF): Values < 1.0x indicate faster than real-time synthesis (e.g. 0.25x = 4× real-time speed; 0.03x = 33× real-time speed). System SAPI and Piper achieve immediate sub-second voice response suitable for real-time conversational agents.

API Quick Reference

Method / Signature	Return Type	Description	Docs
`FastTTS.standard()`	`FastTTS`	Constructs orchestrator with all auto-discovered local and cloud engines pre-registered.	Reference
`registerBackend(FastTTSBackend backend)`	`void`	Registers a custom TTS backend with the orchestrator.	Reference
`speak(String text)`	`FastTTSAudio`	Synthesizes text to audio using the active backend.	Reference
`speak(String backend, String text, FastTTSVoice voice, FastTTSConfig config)`	`FastTTSAudio`	Synthesizes text with specific backend and configuration.	Reference
`prepare()` / `prepareAll()`	`void`	Pre-warms model weights and runtimes into OS page cache.	Reference
`close()`	`void`	Closes and releases all held native subprocesses, threads, and handles.	Reference
`stream(backend, text, voice, config, consumer)`	`void`	Streams audio chunks directly to a byte consumer callback.	Reference
`use(String backendName)`	`void`	Sets the active backend by name.	Reference
`getAllVoices()`	`List<FastTTSVoice>`	Returns available voices across all registered backends.	Reference

Technical Demos & Benchmarks

Case	Java Example	Launcher	Description
Multi-Backend CLI Demo	Demo.java	`run-demo.bat`	Command-line speech synthesis across Windows SAPI, Piper, ElevenLabs, and Deepgram.
Engine & Model Installer	FastTTSInstaller.java	`run-installer.bat`	Interactive engine installer & model manager with `--install` and `--deinstall`.
JMH Microbenchmark Suite	Benchmark.java	`run-benchmark.bat`	Formal OpenJDK JMH throughput and latency benchmarks for TTS synthesis.

Installation

Option 1: Maven (Recommended)

Add the JitPack repository and the complete dependency stack to your pom.xml:

<repositories>
    <repository>
        <id>jitpack.io</id>
        <url>https://jitpack.io</url>
    </repository>
</repositories>

<dependencies>
    <!-- FastTTS Engine -->
    <dependency>
        <groupId>com.github.andrestubbe</groupId>
        <artifactId>FastTTS</artifactId>
        <version>0.1.6</version>
    </dependency>

    <!-- FastSIMD Hardware Vector Acceleration Engine -->
    <dependency>
        <groupId>com.github.andrestubbe</groupId>
        <artifactId>FastSIMD</artifactId>
        <version>0.1.3</version>
    </dependency>

    <!-- FastMemory Aligned Allocator -->
    <dependency>
        <groupId>com.github.andrestubbe</groupId>
        <artifactId>FastMemory</artifactId>
        <version>0.1.1</version>
    </dependency>

    <!-- FastPointer Address Wrapper -->
    <dependency>
        <groupId>com.github.andrestubbe</groupId>
        <artifactId>FastPointer</artifactId>
        <version>0.1.1</version>
    </dependency>

    <!-- FastAudioProcess Audio Engine -->
    <dependency>
        <groupId>com.github.andrestubbe</groupId>
        <artifactId>FastAudioProcess</artifactId>
        <version>0.1.1</version>
    </dependency>

    <!-- FastCore Unified JNI Loader -->
    <dependency>
        <groupId>com.github.andrestubbe</groupId>
        <artifactId>FastCore</artifactId>
        <version>0.1.1</version>
    </dependency>
</dependencies>

Option 2: Gradle (via JitPack)

repositories {
    maven { url 'https://jitpack.io' }
}

dependencies {
    implementation 'com.github.andrestubbe:FastTTS:0.1.6'
    implementation 'com.github.andrestubbe:FastSIMD:0.1.3'
    implementation 'com.github.andrestubbe:FastMemory:0.1.1'
    implementation 'com.github.andrestubbe:FastPointer:0.1.1'
    implementation 'com.github.andrestubbe:FastAudioProcess:0.1.1'
    implementation 'com.github.andrestubbe:FastCore:0.1.1'
}

Engine & Model Manager (CLI Installer)

After adding FastTTS to your project or cloning the repository, use the bundled CLI installer to manage offline neural runtimes and models with automated zero-config setup:

# Launch interactive console dashboard:
run-installer.bat

# Check engine and model status:
run-installer.bat --status

# Install all standard offline engines & models (Kokoro, Pocket TTS, Piper, Matcha, VITS):
run-installer.bat --install all

# Install or deinstall specific engines:
run-installer.bat --install matcha
run-installer.bat --install vits
run-installer.bat --install kokoro
run-installer.bat --install pocket
run-installer.bat --install piper
run-installer.bat --install f5tts
run-installer.bat --install dia2
run-installer.bat --deinstall piper

Supported Installer Actions:

--status — Displays live tabular matrix of all offline model weights and binaries.
--install <all|kokoro|matcha|vits|pocket|piper|f5tts|omnivoice|dia2> — Automatically fetches official releases, verifies integrity, extracts archives (.tar.bz2, .zip), and registers model weights locally.
--deinstall <all|kokoro|matcha|vits|pocket|piper|f5tts|omnivoice|dia2> — Cleanly purges local models and binaries for reproducible resets.

Backend Setup & Configuration

1. Offline Neural & System Backends (Zero Manual Setup)

All local engines run 100% offline with zero Python dependencies. FastTTS auto-resolves models and runtimes in ./bin and ./models (or ~/.speak/). Use the bundled CLI installer to fetch everything automatically:

# One-command installation for all standard local offline engines:
run-installer.bat --install all

Windows SAPI — Pre-installed on Windows 10/11 (fasttts.dll via JNI, ~138 ms latency). Zero setup required.
Kokoro Studio — Sherpa-ONNX runtime + 10 studio voices (af_bella, am_adam, bf_emma, etc.) at 24 kHz. Installed via run-installer.bat --install kokoro.
Matcha-TTS — Optimal-Transport Flow Matching neural TTS (matcha-icefall + Vocos vocoder at 22.05 kHz). Over 10x real-time speed on CPU (~0.09x RTF). Installed via run-installer.bat --install matcha.
VITS (Piper ONNX) — End-to-end variational inference for low-latency deterministic neural speech (~0.09x RTF). Installed via run-installer.bat --install vits.
Pocket TTS — High-speed quantized INT8 neural CPU inference (~0.71x RTF). Installed via run-installer.bat --install pocket.
Piper Neural — Lightweight offline ONNX engine with German & English voices. Installed via run-installer.bat --install piper.
F5-TTS — Non-autoregressive Flow Matching voice cloning & TTS via audiocpp_cli.exe. Installed via run-installer.bat --install f5tts.
Dia2 Dialogue — Multi-speaker scripted dialogue ([S1], [S2]) via audiocpp_cli.exe and GGUF 2B weights. Installed via run-installer.bat --install dia2.
OmniVoice — Multilingual diffusion zero-shot voice cloning (600+ languages) with reference audio prompts. Installed via run-installer.bat --install omnivoice.

2. Cloud Backends (API Credentials)

Cloud backends stream real-time high-fidelity speech over HTTPS/WebSocket. Simply set the corresponding environment variables:

Cloud Backend	Environment Variable	Dashboard / Token Registration
Hugging Face	`HF_TOKEN`	Hugging Face Access Tokens
ElevenLabs	`ELEVENLABS_API_KEY`	ElevenLabs Dashboard
Deepgram	`DEEPGRAM_API_KEY`	Deepgram Console
OpenAI	`OPENAI_API_KEY`	OpenAI Platform

# Example (Windows CMD):
set HF_TOKEN=hf_your_token_here
set ELEVENLABS_API_KEY=your_key_here

Documentation

CHANGELOG.md: Release notes and version history.
REFERENCE.md: Core API reference manual.
PHILOSOPHY.md: Engineering rationale for zero-allocation performance.
COMPILE.md: Full compilation guide (MSVC C++17 build chain + JNI Setup).
ROADMAP.md: Future development goals.

Platform Support

Platform	Architecture	Status	Notes
Windows 10/11	x64	✅ Fully Supported	Native Windows SAPI and MSVC AVX2 compilation
Linux	x64, ARM64	🚧 Planned	Native Piper and ALSA / Pulse audio support
macOS	Apple Silicon, x64	🚧 Planned	AVFoundation / Piper support

License

MIT License — See LICENSE file for details.

Related Projects

FastCore — Native library loader, FFM gateway, and platform abstraction
FastAudioCapture — Zero-copy WASAPI audio loopback & microphone capture
FastAudioPlayer — Ultra low-latency WASAPI and DirectSound audio playback engine
FastAudioProcess — Zero-allocation audio DSP, resamplers, and format converters
FastVAD — Ultra-fast real-time Voice Activity Detection (Silero-ONNX & WebRTC)
FastTurn — Real-time conversational turn-taking and speech arbitration
FastSSML — High-performance SSML parsing and speech synthesis markup
FastSTT — High-throughput local speech-to-text recognition
FastWakeWord — Neural wake-word and keyword spotting detector

Part of the FastJava Ecosystem — Making the JVM faster. Small package. Maximum speed. Zero bloat. 🚀📋

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
docs		docs
examples		examples
native		native
release		release
src/main		src/main
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
TODO.txt		TODO.txt
compile.bat		compile.bat
jitpack.yml		jitpack.yml
pom.xml		pom.xml
run-benchmark.bat		run-benchmark.bat
run-demo.bat		run-demo.bat
run-installer.bat		run-installer.bat

Navigation Menu

Repository navigation

About

Repository files navigation

FastTTS 0.1.6 [ALPHA-2026-10-17] — Unified, Zero-Bloat TTS Backend Orchestration for Java

Quick Start — Example

Table of Contents

Why FastTTS?

Key Features

Real-World Use Cases

Architecture Overview

Core Backend Engines

Performance Benchmarks

API Quick Reference

Technical Demos & Benchmarks

Installation

Option 1: Maven (Recommended)

Option 2: Gradle (via JitPack)

Engine & Model Manager (CLI Installer)

Supported Installer Actions:

Backend Setup & Configuration

1. Offline Neural & System Backends (Zero Manual Setup)

2. Cloud Backends (API Credentials)

Documentation

Platform Support

License

Related Projects

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

FastTTS 0.1.6 [ALPHA-2026-10-17] — Unified, Zero-Bloat TTS Backend Orchestration for Java

Quick Start — Example

Table of Contents

Why FastTTS?

Key Features

Real-World Use Cases

Architecture Overview

Core Backend Engines

Performance Benchmarks

API Quick Reference

Technical Demos & Benchmarks

Installation

Option 1: Maven (Recommended)

Option 2: Gradle (via JitPack)

Engine & Model Manager (CLI Installer)

Supported Installer Actions:

Backend Setup & Configuration

1. Offline Neural & System Backends (Zero Manual Setup)

2. Cloud Backends (API Credentials)

Documentation

Platform Support

License

Related Projects

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages