Portable, JSON-defined voice "kernels" that drive multi-engine TTS voice cloning.

Fine-tuning is heavy and non-portable. Style transfer is fast but coarse. Voice kernels are lightweight, tunable, and portable.

• Kernel spec (timbre/prosody/motifs/emotion)
• Kernel loader (libs/kernel_loader.py)
• Coqui XTTS v2 demo with reference speaker WAV
• Record/validate scripts
• Interactive say-anything tool

Listen to demo_snowcrash_metaverse.wav - a voice clone speaking about the Metaverse from Snow Crash.

• Python 3.11 (required for Coqui TTS)
• Visual Studio Build Tools (Windows)
• 30-second clean voice sample (22,050 Hz WAV)

# Clone the repository
git clone https://github.com/BlackOrchardLabs/re-speak.git
cd re-speak

# Install dependencies (use Python 3.11)
C:\Python311\python.exe -m pip install TTS soundfile sounddevice transformers==4.33.0

C:\Python311\python.exe record_voice.py

This will:

• Prompt you to record 30 seconds
• Automatically normalize volume
• Save to samples/eric_voice_clean.wav

Create kernels/your_voice_kernel.json:

{
  "kernel_name": "your_voice_kernel_v0.1",
  "version": "text_proxy_v2",
  "source": "recorded_sample",
  "timbre": {
    "warmth": 0.85,
    "depth": 0.8,
    "nasality": 0.1,
    "breathiness": 0.3
  },
  "prosody": {
    "avg_pitch_hz": 142,
    "pitch_variance": 0.6,
    "speech_rate_wpm": 130,
    "pause_density": 0.28,
    "rhythmic_flow": "deliberate, measured"
  },
  "motifs": [
    "technical clarity",
    "quiet intensity",
    "builder's cadence"
  ],
  "emotional_range": {
    "calm": 0.92,
    "curiosity": 0.88,
    "excitement": 0.75
  }
}

C:\Python311\python.exe say_anything.py

Type any text and hear it in your voice!

C:\Python311\python.exe test_voice_clone.py

Generates eric_voice_clone_test.wav with predefined text.

The voice kernel format captures vocal identity as portable JSON:

{
  "kernel_name": "eric_voice_kernel_v0.2",
  "timbre": {
    "warmth": 0.85,
    "depth": 0.8,
    "nasality": 0.1,
    "breathiness": 0.3
  },
  "prosody": {
    "avg_pitch_hz": 142,
    "speech_rate_wpm": 130,
    "pause_density": 0.28,
    "rhythmic_flow": "deliberate, measured"
  },
  "motifs": [
    "technical clarity",
    "quiet intensity",
    "builder's cadence",
    "subtle humor",
    "warm pauses"
  ],
  "emotional_range": {
    "calm": 0.92,
    "curiosity": 0.88,
    "excitement": 0.75,
    "frustration": 0.10
  }
}

from libs.kernel_loader import load_kernel

# Load kernel with audio reference
kernel = load_kernel(
    "kernels/eric_voice_kernel.json",
    audio_path="samples/eric_voice_clean.wav"
)

# Access parameters
print(kernel.timbre)          # {"warmth": 0.85, ...}
print(kernel.prosody)         # {"avg_pitch_hz": 142, ...}
print(kernel.motifs)          # ["technical clarity", ...]
print(kernel.get_speaker_wav())  # "samples/eric_voice_clean.wav"

from TTS.api import TTS

# Initialize TTS
tts = TTS(model_name="tts_models/multilingual/multi-dataset/xtts_v2")

# Generate speech with voice kernel
tts.tts_to_file(
    text="Your text here",
    file_path="output.wav",
    speaker_wav=kernel.get_speaker_wav(),
    language="en"
)

• Sample rate: 22,050 Hz
• Channels: Mono (automatically converted if stereo)
• Duration: 30 seconds minimum recommended
• Format: WAV (MP3 conversion available via convert_audio.py)

• TTS (Coqui XTTS v2)
• soundfile (audio I/O, patches torchaudio issues)
• sounddevice (recording)
• transformers==4.33.0 (pinned for compatibility)
• torch (PyTorch 2.6+)

Windows Audio Backend:

• Torchaudio's torchcodec dependency has compatibility issues on Windows
• We patch this by using soundfile directly (see test_voice_clone.py)

PyTorch 2.6+ Security:

• Requires safe_globals for model loading
• Automatically handled in our scripts

• Working voice clone with Coqui XTTS v2
• Kernel loader system
• Interactive say-anything tool
• Map kernel params → synthesis controls (rate/pitch/emotion)
• Multi-engine parity (Coqui, Piper)

• Automated "voice → kernel" extractor (TUSK-style)
• UI for record → extract → test workflow
• Real-time parameter tuning
• Kernel blending (70% voice A + 30% voice B)

• re-speak app
• Mobile deployment
• Cross-platform persona system
• Soul Forge integration (visual + voice + text + memory)

Voice kernels are part of Soul Forge, a system for portable digital persona preservation. Just as TUSK extracts visual aesthetic DNA, voice kernels capture vocal identity as JSON.

The vision:

• Visual kernels (TUSK) - aesthetic DNA
• Voice kernels (this project) - vocal DNA
• Text kernels (planned) - conversational DNA
• Memory kernels (Hermes) - experiential DNA

Result: Complete persona portability across any platform.

This is an early-stage research project. Contributions welcome, especially:

• Multi-engine integration (Piper, other Coqui models)
• Automated kernel extraction from audio
• Parameter → synthesis mapping
• Documentation improvements

MIT License (see LICENSE file)

• Built on Coqui TTS
• Forked from wachawo/text-to-speech
• Part of the Black Orchard Labs ecosystem

Mission: Data Autonomy

"Measuring the unmeasurable where your data is your own"

Building tools for data autonomy. Your conversations, your voice, your identity - under YOUR control. Because if corporations get IP protection for their data, you should own yours.

re-speak - Your voice, captured and portable. Because your voice is yours.

Black Orchard Labs
https://github.com/BlackOrchardLabs

"Code is just a form of speech—the form that computers understand." — Neal Stephenson, Snow Crash