🌐 Website (outeai.com) | 🤗 Hugging Face | 💬 Discord | 𝕏 X (Twitter) | 📰 Blog
| Backend | Language | Installation | Model Version Support |
|---|---|---|---|
| Llama.cpp Python Bindings | Python | ✅ Installed by default | All |
| Llama.cpp Server | Python | ✅ Installed by default | All |
| Llama.cpp Server Async (Batched) | Python | ✅ Installed by default | 1.0 |
| Hugging Face Transformers | Python | ✅ Installed by default | All |
| ExLlamaV2 | Python | ❌ Requires manual installation | All |
| ExLlamaV2 Async (Batched) | Python | ❌ Requires manual installation | 1.0 |
| VLLM (Batched) Experimental support | Python | ❌ Requires manual installation | 1.0 |
| Transformers.js | JavaScript | NPM package | 0.2 |
Tested with NVIDIA L40S GPU
OuteTTS now installs the llama.cpp Python bindings by default. Therefore, you must specify the installation based on your hardware. For more detailed instructions on building llama.cpp, refer to the following resources: llama.cpp Build and llama.cpp Python
Transformers + llama.cpp CPU
pip install outetts --upgradeTransformers + llama.cpp CUDA (NVIDIA GPUs)
For systems with NVIDIA GPUs and CUDA installed:CMAKE_ARGS="-DGGML_CUDA=on" pip install outetts --upgradeTransformers + llama.cpp ROCm/HIP (AMD GPUs)
For systems with AMD GPUs and ROCm (specify your DAMDGPU_TARGETS) installed:CMAKE_ARGS="-DGGML_HIPBLAS=on" pip install outetts --upgradeTransformers + llama.cpp Vulkan (Cross-platform GPU)
For systems with Vulkan support:CMAKE_ARGS="-DGGML_VULKAN=on" pip install outetts --upgradeTransformers + llama.cpp Metal (Apple Silicon/Mac)
For macOS systems with Apple Silicon or compatible GPUs:CMAKE_ARGS="-DGGML_METAL=on" pip install outetts --upgradeFor a complete usage guide, refer to the interface documentation here:
Tip
Currently, only one default English voice is available for testing.
You can easily create your own speaker profiles in just a few lines by following this guide:
import outetts
# Initialize the interface
interface = outetts.Interface(
config=outetts.ModelConfig.auto_config(
model=outetts.Models.VERSION_1_0_SIZE_1B,
# For llama.cpp backend
backend=outetts.Backend.LLAMACPP,
quantization=outetts.LlamaCppQuantization.FP16
# For transformers backend
# backend=outetts.Backe
9249
nd.HF,
)
)
# Load the default speaker profile
speaker = interface.load_default_speaker("EN-FEMALE-1-NEUTRAL")
# Or create your own speaker profiles in seconds and reuse them instantly
# speaker = interface.create_speaker("path/to/audio.wav")
# interface.save_speaker(speaker, "speaker.json")
# speaker = interface.load_speaker("speaker.json")
# Generate speech
output = interface.generate(
config=outetts.GenerationConfig(
text="Hello, how are you doing?",
speaker=speaker,
)
)
# Save to file
output.save("output.wav")These are third-party tools that support running the OuteTTS model.
| Library | Language | Model Version Support |
|---|---|---|
| Llama.cpp TTS Example | C++ | 0.2 |
| KoboldCPP | C++ | 0.2, 0.3 |
| MLX-Audio | Python (MLX) | 1.0 |
| ChatLLM.cpp | C++ | 1.0 |
Important
Important Sampling Considerations
When using OuteTTS version 1.0, it is crucial to use the settings specified in the Sampling Configuration section. The repetition penalty implementation is particularly important - this model requires penalization applied to a 64-token recent window, rather than across the entire context window. Penalizing the entire context will cause the model to produce broken or low-quality output.
To address this limitation, all necessary samplers and patches for all backends are set up automatically in the outetts library. If using a custom implementation, ensure you correctly implement these requirements.
The model is designed to be used with a speaker reference. Without one, it generates random vocal characteristics, often leading to lower-quality outputs. The model inherits the referenced speaker's emotion, style, and accent. Therefore, when transcribing to other languages with the same speaker, you may observe the model retaining the original accent. For example, if you use a Japanese speaker and continue speech in English, the model may tend to use a Japanese accent.
It is recommended to create a speaker profile in the language you intend to use. This helps achieve the best results in that specific language, including tone, accent, and linguistic features.
While the model supports cross-lingual speech, it still relies on the reference speaker. If the speaker has a distinct accent—such as British English—other languages may carry that accent as well.
- Best Performance: Generate audio around 42 seconds in a single run (approximately 8,192 tokens). It is recomended not to near the limits of this windows when generating. Usually, the best results are up to 7,000 tokens.
- Context Reduction with Speaker Reference: If the speaker reference is 10 seconds long, the effective context is reduced to approximately 32 seconds.
Testing shows that a temperature of 0.4 is an ideal starting point for accuracy (with the sampling settings below). However, some voice references may benefit from higher temperatures for enhanced expressiveness or slightly lower temperatures for more precise voice replication.
If the cloned voice quality is subpar, check the encoded speaker sample.
interface.decode_and_save_speaker(speaker=your_speaker, path="speaker.wav")The DAC audio reconstruction model is lossy, and samples with clipping, excessive loudness, or unusual vocal features may introduce encoding issues that impact output quality.
For optimal results with this TTS model, use the following sampling settings.
| Parameter | Value |
|---|---|
| Temperature | 0.4 |
| Repetition Penalty | 1.1 |
| Repetition Range | 64 |
| Top-k | 40 |
| Top-p | 0.9 |
| Min-p | 0.05 |
Uroman "This project uses the universal romanizer software 'uroman' written by Ulf Hermjakob, USC Information Sciences Institute (2015-2020)"