VibeVoice

Best for: developers who need speech synthesis or transcription running on their own hardware rather than through a paid API.

VibeVoice is Microsoft's open-source frontier voice AI, covering both text-to-speech and, through VibeVoice-ASR, speech recognition. It is MIT licensed, which is unusually permissive for a model of this capability. VibeVoice-ASR is integrated into Azure AI Foundry Labs and included in a Transformers release, so it works with standard tooling.

What makes it notable: the edge inference work. VibeVoice-ASR-BitNet uses heterogeneous quantisation to compress the model from 4.62 GB to 1.58 GB while maintaining real-time inference on edge CPUs — which puts speech recognition on modest hardware without a GPU, and without sending audio to a third party.

Where it falls short: running voice models yourself means handling model weights, dependencies and inference infrastructure, and the full-size models still want a GPU for comfortable throughput. Voice cloning capability carries obvious misuse potential, so consent and disclosure matter. As a research release, documentation is thinner than a commercial API's.

Details: Python, MIT licensed. Backed by Microsoft.