Model Description
OuteTTS-0.1-350M is a novel text-to-speech synthesis model that leverages pure language modeling without external adapters or complex architectures, built upon the LLaMa architecture using our Oute3-350M-DEV base model, it demonstrates that high-quality speech synthesis is achievable through a straightforward approach using crafted prompts and audio tokens.
Key Features
- Pure language modeling approach to TTS
- Voice cloning capabilities
- LLaMa architecture
- Compatible with llama.cpp and GGUF format
Technical Details
The model utilizes a three-step approach to audio processing:
- Audio tokenization using WavTokenizer (processing 75 tokens per second)
- CTC forced alignment for precise word-to-audio token mapping
- Structured prompt creation following the format:
[full transcription]
[word] [duration token] [audio tokens]
Technical Blog
https://www.outeai.com/blog/OuteTTS-0.1-350M
Limitations
Being an experimental v0.1 release, there are some known issues:
- Vocabulary constraints due to training data limitations
- String-only input support
- Given its compact 350M parameter size, the model may frequently alter, insert, or omit wrong words, leading to variations in output quality.
- Variable temperature sensitivity depending on use case
- Performs best with shorter sentences, as accuracy may decrease with longer inputs
Speech Samples
Listen to samples generated by OuteTTS-0.1-350M:
Installation
pip install outetts
Usage
[!WARNING] The example below works with older
outettsversion (==0.1.7). The new version (>=0.2.0) introduces changes to the interface. Please refer to the GitHub Usage Example for updated examples.
Interface Usage
from outetts.v0_1.interface import InterfaceHF, InterfaceGGUF
# Initialize the interface with the Hugging Face model
interface = InterfaceHF("OuteAI/OuteTTS-0.1-350M")
# Or initialize the interface with a GGUF model
# interface = InterfaceGGUF("path/to/model.gguf")
# Generate TTS output
# Without a speaker reference, the model generates speech with random speaker characteristics
output = interface.generate(
text="Hello, am I working?",
temperature=0.1,
repetition_penalty=1.1,
max_length=4096
)
# Play the generated audio
output.play()
# Save the generated audio to a file
output.save("output.wav")
Voice Cloning
# Create a custom speaker from an audio file
speaker = interface.create_speaker(
"path/to/reference.wav",
"reference text matching the audio"
)
# Generate TTS with the custom voice
output = interface.generate(
text="This is a cloned voice speaking",
speaker=speaker,
temperature=0.1,
repetition_penalty=1.1,
max_length=4096
)
Model Details
- Model Type: LLaMa-based language model
- Size: 350M parameters
- Language Support: English
- License: CC BY 4.0
- Speech Datasets Used:
- LibriTTS-R (CC BY 4.0)
- Multilingual LibriSpeech (MLS) (CC BY 4.0)
Future Improvements
- Scaling up parameters and training data
- Exploring alternative alignment methods for better character compatibility
- Potential expansion into speech-to-speech assistant models
Credits
- WavTokenizer: https://github.com/jishengpeng/WavTokenizer
- CTC Forced Alignment: https://pytorch.org/audio/stable/tutorials/ctc_forced_alignment_api_tutorial.html
Disclaimer
By using this model, you acknowledge that you understand and assume the risks associated with its use. You are solely responsible for ensuring compliance with all applicable laws and regulations. We disclaim any liability for problems arising from the use of this open-source model, including but not limited to direct, indirect, incidental, consequential, or punitive damages. We make no warranties, express or implied, regarding the model's performance, accuracy, or fitness for a particular purpose. Your use of this model is at your own risk, and you agree to hold harmless and indemnify us, our affiliates, and our contributors from any claims, damages, or expenses arising from your use of the model.