MOSS-TTS Unified Interface

A unified web interface combining all MOSS-TTS models into a single application with an intuitive tabbed interface.

Features

Voice Synthesis Models

• MOSS-TTS-v1.5 - High-fidelity voice cloning (31 languages, pause markers, improved cloning)
• MOSS-TTSD - Multi-speaker dialogue generation
• MOSS-VoiceGenerator - Design voices from text descriptions
• MOSS-SoundEffect - Generate environmental sounds and effects
• MOSS-TTS-Realtime - Low-latency streaming TTS for voice agents

Interface

• Single unified interface with tab-based navigation
• Smart on-demand model loading to optimize memory usage
• Modern, responsive UI built with Gradio

Quick Start with Pinokio

This application is packaged for Pinokio for one-click installation and management.

Available Commands:

• Install - Sets up Python environment, installs dependencies, and configures PyTorch for your GPU
• Start - Launches the Gradio UI on 127.0.0.1 using Pinokio’s next available port (avoiding conflicts). After startup, use Open Web UI in Pinokio or the URL printed in the terminal.
• Update - Pulls latest changes from repository
• Reset - Removes the env virtual environment for a clean reinstall

Programmatic access

• Pinokio (launcher scripts) — Run actions from the Pinokio UI, or invoke scripts programmatically with Pinokio’s script.start API (for example from another launcher script) using the target script’s file name as uri and any params your flow needs. Launcher scripts live in the project root (install.js, start.js, update.js, reset.js, link.js).
• Python — Application code lives under app/. After dependencies are installed (virtualenv at project root), from the repo root:
  python app/app.py --host 127.0.0.1 --port <port>
  Or: cd app then python app.py .... Defaults in app/app.py are --host (platform-specific) and --port 7860 if you omit flags.
• HTTP (curl) — The Gradio server is a normal HTTP app. Once it is listening, you can probe it with curl, for example:
  curl -sS -I http://127.0.0.1:<port>/
  Replace <port> with the port shown at startup (Pinokio assigns a free port when you start from the launcher).

System Requirements

Minimum:

• Python 3.10 or higher (3.12 recommended)
• 16GB RAM
• 50GB free disk space
• Internet connection for model downloads

Recommended:

• 32GB RAM
• NVIDIA GPU with 10GB+ VRAM (24GB+ for optimal performance)
• CUDA 12.8 compatible drivers

Note: CPU-only mode is supported but significantly slower.

Usage Guide

Voice Cloning (TTS)

Generate speech with optional voice cloning from reference audio (default: MOSS-TTS-v1.5).

1. Enter your text (use [pause 3.2s] in text for explicit pauses on v1.5)
2. For non Chinese/English, pick a Language tag when using the 8B model
3. Optionally upload reference audio (3-30 seconds recommended)
4. Adjust generation settings if needed
5. Click "Generate Speech"

Without reference audio: Uses default voice
With reference audio: Clones the voice characteristics

Dialogue Generation (TTSD)

Create multi-speaker conversations with distinct voices.

1. Write your dialogue with speaker tags:

[S1] Hello there!
[S2] Hi! How are you?
[S1] Great weather today.

2. Set the number of speakers (1-5)
3. Click "Generate Dialogue"

Voice Design (VoiceGenerator)

Create custom voices from text descriptions without reference audio.

1. Describe the desired voice characteristics:
   • Age and gender
   • Tone and emotion
   • Accent or style
2. Enter text to synthesize
3. Click "Generate Voice"

Example descriptions:

• "A young female with a cheerful, energetic tone"
• "An elderly male with a calm, wise voice"
• "A middle-aged professional with a confident tone"

Sound Effects

Generate environmental sounds and audio effects from descriptions.

1. Describe the sound you want:
   • "Thunder and rain in a storm"
   • "Busy city street with traffic"
   • "Crackling fireplace"
2. Click "Generate Sound"

Generation Settings

• Temperature (0.1-3.0): Controls randomness. Lower = more stable, higher = more creative
• Top P (0.1-1.0): Nucleus sampling threshold for token selection
• Top K (1-200): Limits vocabulary selection to top K tokens
• Max New Tokens: Controls maximum output length

Memory Usage

Figures are weights only; generation adds a KV cache that grows with max_new_tokens.

Only one model stays resident in GPU memory at a time. When you switch tabs (or generate with a different model) the previous model's VRAM is freed before the new one loads, so a single 24GB card is plenty for any individual model. If you exceed VRAM, the NVIDIA driver silently spills to system RAM and generation becomes extremely slow — keeping one model resident avoids that.

Tuning memory & speed

• --quantization auto (the Pinokio default) loads 4-bit weights on CUDA cards with ≤32GB VRAM when bitsandbytes is installed, and falls back to bf16 otherwise. This keeps the 8B SFX/Dialogue models around ~6GB so they never overflow a 24GB card.
• --quantization 4bit forces the same 4-bit path (needs pip install bitsandbytes, CUDA only). Use none to force full bf16.
• --model_cache_size trades VRAM for speed: raise it on a large-VRAM card to avoid reloading a model each time you revisit its tab; keep it at 1 on a 24GB card.

Troubleshooting

Out of Memory / very slow generation (especially SFX & Dialogue)

• 4-bit quantization is on by default (auto) so the 8B models fit in ~6GB. If generation is still slow, the most common cause on Windows is below.
• On Windows, disable the NVIDIA driver's CUDA - System Memory Fallback: NVIDIA Control Panel → Manage 3D Settings → CUDA - Sysmem Fallback Policy → Prefer No Sysmem Fallback. When VRAM is exceeded this driver feature silently spills to system RAM over PCIe, making generation 10–100× slower (the classic "fills VRAM and becomes glacial" symptom) instead of erroring.
• Confirm bitsandbytes installed (auto falls back to bf16 without it — check the startup log for the [quantization] auto: line)
• Close other GPU applications
• Reduce max_new_tokens / target duration
• Use CPU mode if GPU memory is insufficient

Installation Issues

• Run Reset in Pinokio
• Run Install again
• Check Python version compatibility

Model Download Failures

• Verify internet connection
• Ensure sufficient disk space (~50GB)
• Check firewall settings

Poor Audio Quality

• Use high-quality reference audio (clear, minimal background noise)
• Adjust temperature setting (try 0.7-1.0 range)
• Ensure reference audio is 3-30 seconds long

Best Practices

1. Test incrementally - Start with short text to verify settings
2. Quality reference audio - Use clear recordings with minimal background noise
3. Descriptive prompts - Provide detailed descriptions for voice/sound generation
4. Adjust settings - Experiment with temperature and sampling parameters
5. Monitor memory - Close unused applications when running large models

Resources

• Source Code: OpenMOSS/MOSS-TTS
• Issues: GitHub Issues
• Community: OpenMOSS Discord

License

Apache 2.0 License (same as MOSS-TTS)

Acknowledgments

• OpenMOSS Team - MOSS-TTS model development
• MOSI.AI - Research and model training
• Gradio - Web interface framework