Microsoft drops VibeVoice for long-form audio synthesis

VibeVoice is a major milestone for open-source speech AI, offering a locally runnable alternative to proprietary giants like ElevenLabs.

• –3200x audio compression via novel continuous speech tokenizers enables long-form generation on hardware with as little as 8GB VRAM.
• –Native support for four distinct speakers with natural turn-taking simplifies the automation of podcasting and meeting transcription workflows.
• –While internal benchmarks claim parity with ElevenLabs V3, real-world users report occasional stability issues and "hallucinated" background artifacts.
• –The model's cross-lingual capabilities and low-latency realtime variant (0.5B) make it highly versatile for interactive agent applications.
• –Microsoft's decision to restrict TTS source code shortly after release highlights the ongoing friction between open research and deepfake safety concerns.