Structured Output Benchmark vs VibeVoice

Interfaze's Structured Output Benchmark (SOB) exposes a gap that has been quietly breaking production AI pipelines: models can produce syntactically valid JSON while getting the actual values wrong. SOB measures value accuracy across 21 models using 5,000 text passages, 209 OCR documents, and 115 meeting transcripts — scoring each on seven metrics including value accuracy, faithfulness (grounding vs. hallucination), type safety, and perfect-response rate. The benchmark reveals some sobering findings. Even top models like GPT-5.4 and Claude Sonnet 4.6 achieve ~83% on text but drop to 67% on images and only 23.7% on audio. No single model dominates all modalities — GPT-5.4, GLM-4.7, Qwen3.5-35B, and Gemini 2.5 Flash cluster within one point of each other on text. Perfect response rates (all seven metrics correct) rarely exceed 50% for even the best performers. For developers building data extraction pipelines, agents that read invoices, or any system where "correct JSON" means more than syntactically valid JSON, this is required reading. The dataset is on Hugging Face, the paper is on arXiv, and the playground lets you test your own model's structured output capability directly.

VibeVoice is Microsoft's open-source family of frontier voice AI models covering both speech recognition and synthesis at a scale most commercial services still can't match. The ASR model processes up to 60 minutes of audio in a single pass, generating speaker-diarized, timestamped transcriptions across 50+ languages — complete with hotword customization for domain-specific accuracy. At 7B parameters, it supports on-premise deployment for privacy-sensitive applications. The TTS side is equally impressive: VibeVoice-1.5B synthesizes up to 90 minutes of multi-speaker audio with natural conversational flow and turn-taking between up to four distinct speakers. A lightweight 500M realtime variant streams at under 300ms latency. All of this runs on a novel continuous speech tokenizer operating at just 7.5 Hz — dramatically more efficient than typical audio codecs. What makes this notable is the MIT license. Microsoft isn't just open-sourcing a research demo; they're releasing production-grade weights on Hugging Face alongside code that teams can self-host, fine-tune, or build into their products. With 42,000+ GitHub stars and 771 earned today alone, it's the kind of drop that resets the baseline for what open-source audio AI looks like.