Sopro is a lightweight English text-to-speech model (169M parameters) that aims for decent quality while staying practical to run and tinker with. Instead of a fully Transformer-based stack, it uses dilated convolutions (WaveNet-style) plus lightweight cross-attention, and it’s built around a “streaming” synthesis workflow with optional zero-shot voice cloning (using a short reference clip).
The author reports ~0.25 real-time factor on CPU (on an M3 base model), and suggests using ~3–12 seconds of reference audio for voice cloning. The documentation is unusually helpful for a side project: you can pip install sopro, use the soprotts CLI, or call SoproTTS.from_pretrained() in Python. It’s also candid about limitations: voice similarity depends heavily on recording quality, the streaming and non-streaming outputs won’t match bit-for-bit, and the author notes generation is effectively limited to ~32 seconds before the model starts to hallucinate. If you’re looking for a small, readable TTS codebase to experiment with streaming synthesis, voice conditioning, or model structure tradeoffs (especially on non-GPU hardware), this is a solid starting point.
Quick stats from the listing feed: pipeline: text-to-speech · 20 likes · 280 downloads.
Source listing: https://huggingface.co/models?sort=modified