QORA-TTS-12Hz-0.6B

QORA-TTS 0.6B - Pure Rust Text-to-Speech

Pure Rust TTS engine with 9 built-in speakers. No Python, no CUDA, no external ML frameworks. Single executable + model weights = portable text-to-speech that runs on any machine.

Smart system awareness — automatically detects your hardware (RAM, CPU threads) and adjusts generation limits so TTS runs well even on constrained systems. 9 built-in voices — works out of the box with no reference audio needed. 10 languages supported.

Based on Qwen3-TTS-12Hz-0.6B-CustomVoice (Apache 2.0).

License

This project is licensed under Apache 2.0. The base model Qwen3-TTS-12Hz-0.6B-CustomVoice is released by the Qwen team under Apache 2.0.

What It Does

QORA-TTS 0.6B converts text to natural-sounding speech. It can:

• 9 built-in voices — male and female speakers embedded in the model, no reference audio needed
• 10 languages — English, Chinese, German, Italian, Portuguese, Spanish, Japanese, Korean, French, Russian
• 24 kHz output — high-quality mono WAV audio
• Fast generation — lighter model for quicker speech synthesis
• Controllable generation — adjust length, temperature, and sampling parameters

Quick Start

1. Download all files into the same folder
2. Run:

# Use built-in speaker
qora-tts.exe --speaker ryan --language english --text "Hello, how are you?"

# Different speaker
qora-tts.exe --speaker serena --language chinese --text "你好世界"

# Japanese speaker
qora-tts.exe --speaker ono_anna --language japanese --text "こんにちは"

# Control length and output
qora-tts.exe --speaker aiden --language english --text "Good morning!" --max-codes 200 --output greeting.wav

# Reproducible output with seed
qora-tts.exe --speaker ryan --text "Same every time" --seed 42

Files

qora-tts.exe          4.1 MB   Inference engine
model.qora-tts       971 MB    Q4 weights (talker + predictor + decoder)
config.json           4.8 KB   Model configuration
tokenizer.json         11 MB   Tokenizer (151,936 vocab)
vocab.json            2.7 MB   Vocabulary
merges.txt            1.6 MB   BPE merges
tokenizer_config.json 7.2 KB   Tokenizer config

No safetensors needed. Everything loads from model.qora-tts. The exe auto-finds all files in its own directory.

Architecture

How It Works

1. Text encoding — tokenize input text with 151K BPE vocabulary
2. Speaker selection — load built-in speaker embedding by name
3. Code generation — 28-layer transformer (Talker) generates speech codes autoregressively
4. Code expansion — 5-layer Code Predictor expands code0 into 16 codebooks (codes 0-15)
5. Audio synthesis — VQ decoder + Vocos vocoder converts codes to 24kHz waveform

Smart System Awareness

QORA-TTS detects your system at startup and automatically adjusts generation limits:

QORA-TTS — Pure Rust Text-to-Speech Engine
System: 16384 MB RAM (9856 MB free), 12 threads

Hard caps apply even to explicit user values — if you pass --max-codes 2000 on a system with 6 GB free RAM, it gets clamped to 500 automatically. This prevents the model from running for too long on weak systems.

CLI Arguments

Built-in Speakers

Supported Languages

Performance

Tested on i5-11500 (6C/12T), 16GB RAM, CPU-only:

Example: "Hello, how are you?" (~3 seconds of audio) takes ~10-15 seconds total.

Comparison with 1.7B

Built With

• Language: Pure Rust (2024 edition)
• Dependencies: half (f16), rayon (parallelism), tokenizers (HuggingFace tokenizer), memmap2 (mmap for converter), serde_json (config parsing)
• No ML framework for inference — all matrix ops are hand-written Rust
• Burn framework used only as a build dependency (for binary format types)

Model Binary Format (.qora-tts)

Custom binary format for fast loading:

Header:  "QTTS" magic + version + format byte
Talker:  28 transformer layers (Q4 quantized)
Predictor: 5 transformer layers + code embeddings
Decoder: VQ codebooks + 8 transformer layers + Vocos vocoder