OuteTTS: an experimental text-to-speech model, TTS implemented using a pure language modeling approach

General Introduction

OuteTTS is an experimental text-to-speech (TTS) model that uses a pure language modeling approach to generate high-quality speech. Unlike traditional TTS systems, OuteTTS does not require external adapters or complex architectures. The model is based on the LLaMa architecture and supports a speech cloning feature that enables the generation of speech with random speaker characteristics.OuteTTS aims to achieve efficient speech synthesis through a simple architecture suitable for a wide range of application scenarios.

OuteTTS-0.1-350M is a step forward in simplifying text-to-speech synthesis. OuteTTS-0.1-350M proves that high quality speech can be generated through a purely linguistic modeling approach.

Function List

• text-to-speech: Converts typed text into natural, smooth speech.
• voice cloning: Create custom speakers from reference audio files and generate the corresponding speech.
• Multi-model support: Supports Hugging Face models and GGUF models.
• Audio playback and saving: The generated voice can be played directly or saved as an audio file.
• Temperature and Repeat Penalty: Control the diversity and smoothness of generated speech by adjusting temperature and repetition penalty parameters.

Using Help

Installation process

1. Installing OuteTTS::

   pip install outetts
   

   Important: For GGUF support, you need to manually install the llama-cpp-python. Please visit llama-cpp-python Get specific installation instructions.

Usage

1. Initialize the interface::

   from outetts.v0_1.interface import InterfaceHF, InterfaceGGUF
   # 使用 Hugging Face 模型初始化接口
   interface = InterfaceHF("OuteAI/OuteTTS-0.1-350M")
   # 或者使用 GGUF 模型初始化接口
   # interface = InterfaceGGUF("path/to/model.gguf")
   

2. Generate TTS output::

   output = interface.generate(
   text="Hello, am I working?",
   temperature=0.1,
   repetition_penalty=1.1,
   max_length=4096
   )
   

3. Play and save generated audio::

   # 播放生成的音频
   output.play()
   # 保存生成的音频到文件
   output.save("output.wav")
   

voice cloning

1. Creating custom speakers::

   speaker = interface.create_speaker(
   "path/to/reference.wav",
   "reference text matching the audio"
   )
   

2. Saving and loading speakers::

   # 保存说话人到文件
   interface.save_speaker(speaker, "speaker.pkl")
   # 从文件加载说话人
   speaker = interface.load_speaker("speaker.pkl")
   

3. Generating TTS with Customized Speech::

   output = interface.generate(
   text="This is a cloned voice speaking",
   speaker=speaker,
   temperature=0.1,
   repetition_penalty=1.1,
   max_length=4096
   )
   

parameterization

• Temperature: Controls the diversity of generated speech. Lower temperatures (e.g., 0.1) generate more deterministic outputs, while higher temperatures (e.g., 0.7) generate more diverse outputs.
• Repetition penalty (repetition_penalty): Controls the level of repetition in the generated speech. A higher repetition penalty (e.g., 1.1) reduces the generation of repetitive content.

Through the above steps, users can easily install and use the OuteTTS model for text-to-speech and speech cloning operations. Detailed parameter adjustments and usage examples can help users generate high-quality speech output according to their specific needs.