SoproTTS, a side project, has released version 1.5 of its text-to-speech (TTS) model. This 135M parameter model was trained for approximately $100 using a single GPU.

Performance

SoproTTS v1.5 boasts the following features:

  • 250 ms TTFA streaming latency
  • RTF (Real-Time Factor) of 0.05 (approximately 20× real-time) on CPU
  • Zero-shot voice cloning

The model, while not perfect, represents an improvement over previous versions, offering reduced size, increased speed, and stability. The training code will be made available in the future.

For those evaluating on-premise deployments, there are trade-offs to consider. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these aspects.