OuteTTS 0.3 - Local TTS and Voice Cloning

OuteTTS 0.3 is a local, Apache 2.0–licensed text-to-speech system that also supports voice cloning, letting users generate speech in multiple languages from text and then adapt the voice using an audio sample. The key practical takeaway is that it extends an existing language model (referred to as a base L) with text-to-speech and speech-to-speech capabilities, then fine-tunes that base so it can produce audio directly. In the demo, the model runs on a Google Colab-style setup with a T4 GPU, downloading a relatively small encoder and decoder from the Hugging Face Hub and then using a 1 billion parameter language model backbone.

The model comes in several variants, including a 1B parameter model (also available in GGUF format for use in tools that load LLM-style binaries) and a smaller 500M variant. Language support is a major selling point: the demo highlights English, Japanese, Korean, Chinese, French, and German, with at least one named speaker per language. For English, the available speaker set includes multiple options (one female and three male speakers), and the generation call requires selecting a speaker ID such as an English male option.

Generation is performed by passing user text into a generation configuration that includes common controls like temperature, repetition penalty, and max length, along with the chosen speaker. The notebook workflow is straightforward: import the OuteTTS library (installed as version 0.32 in the environment), download the OuteTTS 0.3 model components, list available speakers, then call a generate method to produce an output audio file (saved as a WAV file in the demo). The demo reports roughly 20 seconds for generation on the T4 GPU, with faster results expected on stronger hardware.

To test quality and controllability, the demo generates English speech for a short lyric-like passage, then increases temperature to make the output sound more expressive. It then switches to the French model using the default French female speaker and produces French audio from a provided sentence.

Voice cloning is presented as an alpha feature. The process starts with an MP3 of a target voice (Dwight Sho in the demo). The system uses Whisper to transcribe the spoken content from that audio into text, then feeds the transcription into OuteTTS with the cloned voice characteristics layered on top. The transcription is described as completely accurate thanks to Whisper, and the resulting cloned voice is used to re-speak the demo introduction text—though the final output is noted as not perfect. Overall, OuteTTS 0.3 positions local TTS and early voice cloning as accessible through a simple notebook workflow, multi-language speaker selection, and configurable generation parameters.