FastVoice 0.1.0 — Ultra-Fast Zero-Shot AI Voice Cloning & Voice Conversion Engine for Java

---

⚡ Hardware-accelerated, zero-shot AI voice cloning and speech-to-speech conversion for the FastJava ecosystem.

FastVoice enables high-fidelity voice cloning in seconds without training heavy neural models. It extracts compact mathematical speaker embeddings (~2 KB) from short reference audio samples (3–10s) and performs instant multi-engine voice cloning and conversion for Java applications.

---

Quick Start

import fastvoice.FastVoice;
import fastvoice.VoiceProfile;
import java.nio.file.Files;
import java.nio.file.Path;

public class Demo {
    public static void main(String[] args) throws Exception {
        // 1. Create a cloned voice profile from a 5s audio sample
        byte[] referenceAudioWav = Files.readAllBytes(Path.of("sample.wav"));
        VoiceProfile profile = FastVoice.createProfile("SpeakerName", "f5tts", referenceAudioWav);

        // 2. Export profile to a target directory
        Path exportDir = Path.of("voices", "clones");
        profile.exportToDirectory(exportDir);
        
        System.out.println("Voice cloned and exported: " + profile.name());
    }
}

---

Table of Contents

• Quick Start
• Why FastVoice?
• Key Features
• Real-World Use Cases
• Architecture Overview
• Performance Benchmarks
• API Quick Reference
• Technical Demos & Benchmarks
• Installation
  • Option 1: Maven (Recommended via JitPack)
  • Option 2: Gradle (via JitPack)
  • Option 3: Direct Download (Pre-built JAR)
• Documentation
• Platform Support
• License
• Related Projects

---

Why FastVoice?

Traditional voice cloning workflows require extensive dataset curation, hours of GPU fine-tuning, and large model distribution packages. FastVoice provides:

• Zero-Shot Speaker Conditioning: Clone timbre from a 3–10s audio sample without fine-tuning weights.
• Ultra-Lightweight Voice Profiles: Acoustic speaker embeddings are stored as compact binary vectors (~2 KB), allowing thousands of voices to be indexed with minimal storage.
• Decoupled Architecture: Clean programmatic Java API separating the voice profile format from the underlying inference engine (F5-TTS, CosyVoice, XTTS).
• Native Ecosystem Integration: Backed by FastCore and designed to pair seamlessly with FastTTS, FastAudioProcess, and FastAudioCapture.

Feature	Legacy Voice Training	Heavy Python Wrappers	FastVoice
Training Requirement	Hours of GPU fine-tuning	Varies	Zero-Shot (3–10s reference audio)
Profile Storage	100 MB – 2 GB per voice	Varies	~2 KB per voice profile
Runtime Overhead	High (PyTorch/CUDA checkpoints)	Heavy process spawning	Pure Java 21+ API & Native FastCore
Voice Conversion (STS)	Separate complex pipelines	External scripts	Built-in FastVoice.convert() API
Downstream Usability	Custom inference scripts	Fragile CLI pipes	Single-line export & synthesis

---

Key Features

• ⚡ Zero-Shot One-Click Cloning: No model training or GPU hours required. Clones timbre from 3–10s of audio in 1–2 seconds.
• 🪶 Ultra-Compact Profiles: Stores embeddings as tiny binary vectors (~2 KB), allowing thousands of voices without disk bloat.
• 🔄 Speech-to-Speech Conversion: Transform spoken recordings into any target profile's voice with prosody preservation.
• 🔌 Seamless Ecosystem Interop: Clean, decoupled Java API designed for integration with FastJava audio and speech pipelines.

---

Real-World Use Cases

• 🎙️ Personalized Voice Assistants: Generate personalized voice responses using the user's or designated speaker's timbre in real time.
• 🎮 Game Character Dialogue & NPCs: Synthesize dynamic speech for dozens of characters using lightweight zero-shot profiles instead of separate multi-gigabyte models.
• 🎬 Video Dubbing & Localization: Convert spoken dialogue into different languages while preserving the original actor's vocal characteristics via Speech-to-Speech conversion.
• 🎧 Accessibility & Voice Preservation: Archive natural voice profiles from short recordings for vocal restoration and text-to-speech assistive devices.

---

Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                 Short Reference Audio (WAV)                 │
│                 (3–10s Clean Voice Sample)                  │
└──────────────────────────────┬──────────────────────────────┘
                               │ FastAudioProcess (DSP Cleaning)
                               ▼
┌─────────────────────────────────────────────────────────────┐
│                          FastVoice                          │
│          • Instant Acoustic Embedding Extraction            │
│          • Zero-Shot Speaker Conditioning                   │
└──────────────────────────────┬──────────────────────────────┘
                               │ Lightweight Profile (~2 KB)
                               ▼
┌─────────────────────────────────────────────────────────────┐
│                     Voice Profile Storage                   │
│        • voice.json  • embedding.bin  • reference.wav       │
└──────────────────────────────┬──────────────────────────────┘
                               │ Programmatic Java API
                               ▼
┌─────────────────────────────────────────────────────────────┐
│                    Downstream Application                   │
│         (Real-time Synthesis & Audio Generation)            │
└─────────────────────────────────────────────────────────────┘

---

Performance Benchmarks

FastVoice is benchmarked using OpenJDK JMH to guarantee sub-microsecond profile instantiation and minimal I/O overhead:

Benchmark / Operation	Throughput (ops/ms)	Ops per Second	Latency
benchmarkProfileCreation (Metadata + VoiceProfile Record)	~147,902 ops/ms	> 147 Million	< 7 ns
benchmarkProfileExport (Disk I/O: JSON + Binary Vector + WAV)	~0.84 ops/ms	~840 ops/s	~1.19 ms

Measured on Windows 11 (x64), JDK 21.0.12.1. Pure Java 21+ record instantiation with zero GC allocation overhead.

---

API Quick Reference

Class / Record	Responsibility	Key Methods	Reference
FastVoice	Central facade for backend registration and synthesis	createProfile(), synthesize(), convert(), registerBackend()	Reference
VoiceProfile	Immutable container for embeddings and reference audio	exportToDirectory(), embedding(), referenceWav()	Reference
VoiceCloner	Interface for cloning engine implementations	extractEmbedding(), cloneAndSynthesize(), convertVoice()	Reference

For complete method signatures and details, refer to docs/REFERENCE.md.

---

Technical Demos & Benchmarks

Case	Java Example	Launcher	Description
Showcase Demo	Demo.java	run-demo.bat	Creates a zero-shot voice profile from reference audio, verifies speaker metadata, and exports it to disk.
JMH Microbenchmark Suite	Benchmark.java	run-benchmark.bat	Formal OpenJDK JMH throughput measurements across profile instantiation and disk export.

---

Installation

Option 1: Maven (Recommended via JitPack)

Add JitPack to your <repositories> and include the dependency in your pom.xml:

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

<dependencies>
    <!-- FastVoice Engine -->
    <dependency>
        <groupId>com.github.andrestubbe</groupId>
        <artifactId>FastVoice</artifactId>
        <version>0.1.0</version>
    </dependency>

    <!-- FastAudioProcess (DSP Denoise & Normalization) -->
    <dependency>
        <groupId>com.github.andrestubbe</groupId>
        <artifactId>FastAudioProcess</artifactId>
        <version>0.1.7</version>
    </dependency>

    <!-- FastCore (Unified Native Platform Loader) -->
    <dependency>
        <groupId>com.github.andrestubbe</groupId>
        <artifactId>FastCore</artifactId>
        <version>0.1.0</version>
    </dependency>
</dependencies>

Option 2: Gradle (via JitPack)

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

dependencies {
    implementation 'com.github.andrestubbe:FastVoice:0.1.0'
    implementation 'com.github.andrestubbe:FastAudioProcess:0.1.7'
    implementation 'com.github.andrestubbe:FastCore:0.1.0'
}

Option 3: Direct Download (Pre-built JAR)

Download the pre-compiled standalone JAR directly from GitHub Releases:

• 📦 FastVoice-0.1.0.jar

---

Documentation

• CHANGELOG.md: Detailed version history and updates.
• REFERENCE.md: Full API documentation and method contracts.
• PHILOSOPHY.md: Zero-shot speaker conditioning design and architecture.
• COMPILE.md: Build instructions and Maven commands.
• ROADMAP.md: Upcoming engine backends and streaming conversions.

---

Platform Support

Platform	Architecture	Status	Notes
Windows 10 / 11	x64	✅ Fully Supported	Native hardware acceleration via FastCore / FastSIMD
Linux	x64 / AArch64	🚧 Planned	Native bindings in development
macOS (Sonoma+)	Apple Silicon / x64	🚧 Planned	Native bindings in development

---

License

MIT License — See LICENSE file for details.

---

Related Projects

• FastTTS — Unified, zero-bloat text-to-speech backend orchestration
• FastSTT — High-throughput local speech-to-text recognition
• FastVAD — Ultra-fast real-time Voice Activity Detection (Silero-ONNX & WebRTC)
• FastAudioCapture — Zero-copy WASAPI audio loopback & microphone capture
• FastAudioProcess — Zero-allocation audio DSP, resamplers, and timeline mixing
• FastAudioPlayer — Ultra low-latency WASAPI and DirectSound audio playback engine
• FastCore — Native library loader, FFM gateway, and platform abstraction
• FastSIMD — Vector hardware acceleration for high-throughput math and DSP

---

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