VocalCore — Open Source Local TTS Platform

VocalCore is a production-quality, open-source local Text-to-Speech (TTS) desktop studio. It is designed to run completely offline on your own hardware, featuring Nvidia/AMD/Intel hardware auto-detection, zero-shot voice cloning, real-time WebSocket audio streaming, training dataset compilers, and model fine-tuning dashboards.

---

⚡ Quick Start (Zero-Setup Launch)

To launch the application:

1. Double-click run.bat at the root of the project folder.
2. The launcher automatically:
   • Sets up a local virtual environment (venv/).
   • Detects your graphics card controller type (NVIDIA, AMD, or Intel).
   • Installs PyTorch optimized for your hardware (CUDA for NVIDIA, DirectML for AMD/Intel, or CPU fallback).
   • Installs all dependencies and builds the React frontend.
   • Starts the FastAPI server and launches the web interface at http://127.0.0.1:7860.

---

💻 Hardware Optimizations (AMD, Intel & NVIDIA)

VocalCore includes native configurations to maximize performance across multi-vendor graphics cards and processor profiles on Windows:

1. GPU Autodetection & DirectML Support:
   • AMD and Intel Arc graphics cards are supported via DirectML (torch-directml). Model tensors are loaded on the DirectML device (privateuseone:0) for hardware-accelerated generation.
2. VRAM Safety Switches:
   • High-end models require substantial memory. The Model Manager automatically unloads the active model from GPU memory and executes PyTorch cache eviction (torch.cuda.empty_cache() / Python garbage collection) before loading a new model to prevent Out-of-Memory (OOM) failures.
   • Users with lower VRAM (e.g. 6GB laptop GPUs) can run lighter models like Qwen3-TTS or Pocket-TTS on the GPU, and toggle CPU Mode in Settings for large models.
3. Half-Precision Acceleration (AMP):
   • Supports BF16 or FP16 precision maps out of the box to cut memory footprint in half and speed up inference.
4. Sub-second WebSocket Streaming:
   • Audio chunks are processed in float buffers and transmitted immediately as binary PCM arrays over WebSockets, enabling playback latency of less than 150ms.

---

📁 Local Model Placement Guide

VocalCore scans local directories for model weights. To avoid automatic HuggingFace downloads and run completely offline, place your weights here:

Model	HuggingFace Repository ID	Local Target Directory
Qwen3-TTS-12Hz (1.7B)	Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice	models/qwen/
Higgs-TTS (3.4B)	bosonai/higgs-tts-3-4b	models/higgs/
Pocket-TTS (100M)	kyutai/pocket-tts	models/pocket/
MOSS-TTS v1.5 (8B)	OpenMOSS-Team/MOSS-TTS-v1.5	models/moss/

Placing Custom Model Files

Inside each folder (e.g. models/qwen/), place:

• config.json and generation_config.json
• model.safetensors or pytorch_model.bin / weights files
• Tokenizers (tokenizer.json, vocab.txt, etc.)

Once the files are present, VocalCore flags them as "Ready" in the web dashboard for instant loading.

---

🛠️ Folder Structure

• run.bat / install.bat / update.bat : Desktop environment loader and compiler scripts
• requirements.txt : Python libraries list
• backend/ : FastAPI routes, streaming hooks, and model adapter loaders
• frontend/ : Vite, React, Lucide Icons, and Canvas audio waveforms
• voices/ : Cloned speaker profiles (JSON metadata + Wav reference audios)
• datasets/ : Compiled speech dataset corpora for fine-tuning
• outputs/ : Generated audio WAVs/MP3s history
• models/ : Offline weights directories

---

🚀 Key Features

• Zero-Shot Voice Cloning: Create voice profiles by uploading a short audio file (minimum 3 seconds, ideal 30-120 seconds).
• Dataset Manager: Slice long recordings on silence boundaries, filter clips based on noise floor SNRs, and transcribe/annotate clips inside the UI.
• Voice Training: Fine-tune custom voice models with play/pause execution loops, real-time loss tracking, ETA countdowns, and TensorBoard loss graphs.
• History Page: Replay, download, filter, or delete previous generations.
• Performance Dashboard: Real-time monitor of CPU, RAM, and VRAM utilization.