MiraTTS (fast 48kHz text-to-speech)
MiraTTS is a text-to-speech model with an unusually performance-focused pitch: the author claims very high throughput (over 100x realtime) when paired with lmdeploy + batching, while still fitting within ~6GB VRAM and keeping latency low. On the quality side, the model targets 48kHz audio output, which can help speech sound less “compressed” than many lower-sample-rate TTS demos.
If you’re building anything interactive (voice UI, agents, real-time narration), the main question is whether those speed claims hold for your deployment shape: single requests vs batched, GPU type, and how much post-processing you need. The quickest way to evaluate is to try the linked Hugging Face Space first to get a feel for voice quality and latency, then follow the GitHub repo instructions to reproduce the runtime setup locally and measure end-to-end performance with your own text lengths and concurrency.
Quick stats from the listing feed: pipeline: text-to-speech · 28 likes · 123 downloads.
Source listing: https://huggingface.co/models?sort=modified